CNTK/DataReader/HTKMLFReader_linux/msra_mgram.h

3170 строки
130 KiB
C++

//
// <copyright file="msra_mgram.h" company="Microsoft">
// Copyright (c) Microsoft Corporation. All rights reserved.
// </copyright>
//
// msra_mgram.h -- simple ARPA LM read and access function
//
#pragma once
#include "basetypes.h"
#include "fileutil.h" // for opening/reading the ARPA file
#include <vector>
#include <string>
#include <hash_map>
#include <algorithm> // for various sort() calls
#include <math.h>
namespace msra { namespace lm {
// ===========================================================================
// core LM interface -- LM scores are accessed through this exclusively
// ===========================================================================
interface ILM // generic interface -- mostly the score() function
{
virtual double score (const int * mgram, int m) const = 0;
virtual bool oov (int w) const = 0; // needed for perplexity calculation
// ... TODO (?): return true/false to indicate whether anything changed.
// Intended as a signal to derived LMs that cache values.
virtual void adapt (const int * data, size_t m) = 0; // (NULL,M) to reset, (!NULL,0) to flush
// iterator for composing models --iterates in increasing order w.r.t. w
interface IIter
{
virtual operator bool() const = 0; // has iterator not yet reached end?
// ... TODO: ensure iterators do not return OOVs w.r.t. user symbol table
// (It needs to be checked which LM type's iterator currently does.)
virtual void operator++() = 0; // advance by one
// ... TODO: change this to key() or something like this
virtual std::pair<const int*,int> operator*() const = 0; // current m-gram (mgram,m)
virtual std::pair<double,double> value() const = 0; // current (logP, logB)
};
virtual IIter * iter (int minM = 0, int maxM = INT_MAX) const = 0;
virtual int order() const = 0; // order, e.g. 3 for trigram
virtual size_t size (int m) const = 0; // return #m-grams
// diagnostics functions -- not all models implement these
virtual int getLastLongestHistoryFound() const = 0;
virtual int getLastLongestMGramFound() const = 0;
};
// ===========================================================================
// log-add helpers
// ===========================================================================
const double logzero = -1e30;
static inline double logadd (double x, double y)
{
double diff = y - x;
double sum = x; // x no longer used after this
if (diff > 0)
{
sum = y; // y no longer used after this
diff = -diff; // that means we need to negate diff
}
if (diff > -24.0) // approx. from a constant from fmpe.h
sum += log (1.0 + exp (diff));
return sum;
}
// take the log, but clip to logzero
template<class FLOATTYPE> // float or double
static inline FLOATTYPE logclip (FLOATTYPE x)
{
// ... TODO: use the proper constants here (slightly inconsistent)
return x > (FLOATTYPE) 1e-30 ? log (x) : (FLOATTYPE) logzero;
}
// compute 1-P in logarithmic representation
static inline double invertlogprob (double logP) { return logclip (1.0 - exp (logP)); }
// ===========================================================================
// CSymbolSet -- a simple symbol table
// ===========================================================================
// compare function to allow char* as keys (without, hash_map will correctly
// compute a hash key from the actual strings, but then compare the pointers
// -- duh!)
struct less_strcmp : public binary_function<const char *, const char *, bool>
{ // this implements operator<
bool operator()(const char * const & _Left, const char * const & _Right) const
{ return strcmp (_Left, _Right) < 0; }
};
class CSymbolSet : public stdext::hash_map<const char *, int, stdext::hash_compare<const char*,less_strcmp>>
{
vector<const char *> symbols; // the symbols
CSymbolSet (const CSymbolSet &); CSymbolSet & operator= (const CSymbolSet &);
public:
CSymbolSet() { symbols.reserve (1000); }
~CSymbolSet() { clear(); }
void clear()
{
foreach_index (i, symbols) free ((void*) symbols[i]);
hash_map::clear();
}
// operator[key] on a 'const' object
// get id for an existing word, returns -1 if not existing
int operator[] (const char * key) const
{
hash_map<const char *,int>::const_iterator iter = find (key);
return (iter != end()) ? iter->second : -1;
}
// operator[key] on a non-'const' object
// determine unique id for a word ('key')
int operator[] (const char * key)
{
hash_map<const char *,int>::const_iterator iter = find (key);
if (iter != end())
return iter->second;
// create
const char * p = _strdup (key);
if (!p)
throw std::bad_exception ("CSymbolSet:id string allocation failure");
try
{
int id = (int) symbols.size();
symbols.push_back (p); // we own the memory--remember to free it
insert (make_pair (p, id));
return id;
}
catch (...)
{
free ((void*) p);
throw;
}
}
// return symbol string for a given id
// Returned pointer is owned by this object.
inline const char * operator[] (int id) const { return symbols[id]; }
// overloads to be compatible with C++ strings and CSymMap
int sym2existingId (const string & key) const { return (*this)[key.c_str()]; }
int sym2id (const string & key) { return (*this)[key.c_str()]; }
inline const char * id2sym (int id) { return (*this)[id]; }
// some helpers for writing and reading back a symbol set
void write (FILE * f)
{
fputTag (f, "SYMS"); // header
fputint (f, (int) size()); // symbol set
foreach_index (k, symbols)
fputstring (f, symbols[k]);
}
void read (FILE * f)
{
clear(); // clear out what was there before (typically nothing)
fcheckTag (f, "SYMS");
int numWords = fgetint (f);
char buf[1000];
for (int k = 0; k < numWords; k++)
{
fgetstring (f, buf);
int id = (*this)[buf];
if (id != k)
RuntimeError ("plsa: sequence error while reading vocabulary");
}
}
};
// ===========================================================================
// mgram_map -- lookup table for mgrams
// ===========================================================================
// variable naming convention for word ids:
// - w a word in user space
// Defined by userSymMap::operator[](string) passed to read().
// Data passed to score() and adapt() functions are in 'w' space.
// - id an id in internal LM space
// E.g. defined by vocabulary in LM input file.
// All external LM accesses involve an implicit mapping, including:
// w -> id --for calls to score() and adapt()
// id -> w --for iterators (IIter++ orders by and *IIter returns keys in 'w' space)
// representation of LM in memory
// LMs are stored sparsely, i.e. only used elements are stored.
// For each m-gram, a score is stored. For each history, a back-off weight is stored.
// Both are stored in flat arrays, one per order, that are concatenations of
// individual arrays per history.
// The mgram_map provides a measure of locating these entries. For each level,
// it stores a flat array of 'firsts' which point to the first child entry in
// the next level (the next 'firsts' value denotes the end).
// The mgram_map also stores word ids, which are the indexes of the sparse
// elements.
// To access an m-gram score of back-off weight, the mgram_map structure is
// traversed, involving a binary search operation at each level.
// a compact vector to hold 24-bit vaulues
class int24_vector : std::vector<unsigned char>
{
public:
// basic (non-tricky) operations --just multiply anything by 3
int24_vector(){}
int24_vector (size_t n) : std::vector<unsigned char> (n*3) {}
void resize (size_t n) { std::vector<unsigned char> & base = *this; base.resize (n*3); }
void reserve (size_t n) { std::vector<unsigned char> & base = *this; base.reserve(n*3); }
void swap (int24_vector & other) { std::vector<unsigned char> & base = *this; base.swap (other); }
size_t size() const { const std::vector<unsigned char> & base = *this; return base.size() / 3; }
bool empty() const { const std::vector<unsigned char> & base = *this; return base.empty(); }
// a reference to a 3-byte int (not a naked pointer as we cannot just assign to it)
template<class T> class uint24_ref_t
{
protected:
T p;
friend class int24_vector; // only int24_vector may instantiate this
__forceinline uint24_ref_t (T p) : p (p) {}
public:
// access
__forceinline operator int () const
{
return (((((signed char) p[2]) << 8) + p[1]) << 8) + p[0];
}
};
typedef uint24_ref_t<const unsigned char *> const_uint24_ref; // const version (only read)
class uint24_ref : public uint24_ref_t<unsigned char *> // non-const (read and assign)
{
static void overflow() { throw runtime_error ("uint32_ref: attempting to store value > 24 bits"); }
protected:
friend class int24_vector; // only int24_vector may instantiate this
__forceinline uint24_ref (unsigned char * p) : uint24_ref_t (p) {}
public:
// assignment operator
__forceinline int operator= (int value)
{
if ((unsigned int) (value+0x800000) > 0xffffff) overflow();
p[0] = (unsigned char) value;
p[1] = (unsigned char) (value >> 8);
p[2] = (unsigned char) (value >> 16);
ASSERT (value == (int) *this);
return value;
}
};
// reading and writing
__forceinline uint24_ref operator[] (size_t i) { std::vector<unsigned char> & base = *this; return uint24_ref (&base[i*3]); }
__forceinline const_uint24_ref operator[] (size_t i) const { const std::vector<unsigned char> & base = *this; return const_uint24_ref (&base[i*3]); }
__forceinline int back() const { const std::vector<unsigned char> & base = *this; return const_uint24_ref (&base[base.size()-3]); }
void push_back (int value)
{
std::vector<unsigned char> & base = *this;
size_t cursize = base.size();
size_t newsize = cursize +3;
if (newsize > base.capacity())
base.reserve (newsize * 2); // double the size to ensure constant-time
base.resize (newsize);
uint24_ref r = uint24_ref (&base[cursize]);
r = value;
ASSERT (value == back());
}
};
// maps from m-grams to m-gram storage locations.
class mgram_map
{
typedef unsigned int index_t; // (-> size_t when we really need it)
//typedef size_t index_t; // (tested once, seems to work)
static const index_t nindex = (index_t) -1; // invalid index
// entry [m][i] is first index of children in level m+1, entry[m][i+1] the end.
int M; // order, e.g. M=3 for trigram
std::vector<std::vector<index_t>> firsts; // [M][i] ([0] = zerogram = root)
std::vector<int24_vector> ids; // [M+1][i] ([0] = not used)
bool level1nonsparse; // true: level[1] can be directly looked up
std::vector<index_t> level1lookup; // id->index for unigram level
static void fail (const char * msg) { throw runtime_error (string ("mgram_map::") + msg); }
// mapping from w -> i -- users pass 'w', internally we use our own 'ids'
std::vector<int> w2id; // w -> id
std::vector<int> id2w; // id -> w
int idmax; // max id ever encountered by create()
inline int map (int w) const
{
if (w < 0 || w >= (int) w2id.size()) return -1;
else return w2id[w];
}
// get index for 'id' in level m+1, as a child of index i in level m.
// Returns -1 if not found.
// This is a relatively generic binary search.
inline index_t find_child (int m, index_t i, int id) const
{
// unigram level is a special case where we can avoid searching
if (m == 0)
{
if (id < 0) return nindex;
index_t i;
if (level1nonsparse)
i = (index_t) id;
else // sparse: use a look-up table
{
if ((size_t) id >= level1lookup.size()) return nindex;
i = level1lookup[id];
}
ASSERT (i == nindex || ids[1][i] == id);
return i;
}
index_t beg = firsts[m][i];
index_t end = firsts[m][i+1];
const int24_vector & ids_m1 = ids[m+1];
while (beg < end)
{
index_t i = (beg + end) / 2;
int v = ids_m1[i];
if (id == v) return i; // found it
else if (id < v) end = i; // id is left of i
else beg = i + 1; // id is right of i
}
return nindex; // not found
}
public:
// --- allocation
mgram_map(){}
mgram_map (int p_M) { init (p_M); }
// construct
void init (int p_M)
{
clear();
M = p_M;
firsts.assign (M, std::vector<index_t> (1, 0));
ids.assign (M+1, int24_vector());
ids[0].resize (1); // fake zerogram entry for consistency
ids[0][0] = -1;
}
// reserve memory for a level
void reserve (int m, size_t size)
{
if (m == 0) return; // cannot reserve level 0
ids[m].reserve (size);
if (m < M)
firsts[m].reserve (size +1);
if (m == 1)
level1lookup.reserve (size);
}
// allow to reduce M after the fact
void resize (int newM)
{
if (newM > M) fail ("resize() can only shrink");
M = newM;
firsts.resize (M);
ids.resize (M+1);
}
// destruct
void clear() { M = 0; firsts.clear(); ids.clear(); w2id.clear(); id2w.clear(); idmax = -1; }
// size
inline int size (int m) const { return (int) ids[m].size(); }
// swap --used e.g. in merging
void swap (mgram_map & other)
{
::swap (M, other.M);
firsts.swap (other.firsts);
ids.swap (other.ids);
::swap (level1nonsparse, other.level1nonsparse);
level1lookup.swap (other.level1lookup);
w2id.swap (other.w2id);
id2w.swap (other.id2w);
::swap (idmax, other.idmax);
}
// --- id mapping
// test whether a word id is known in this model
inline bool oov (int w) const { return map (w) < 0; }
// return largest used word id (=last entry in unigram ids[])
int maxid() const { return idmax; }
// return largest used w (only after created())
int maxw() const { return -1 + (int) w2id.size(); }
// map is indexed with a 'key'.
// A key represents an m-gram by storing a pointer to the original array.
// The key allows to remove predicted word (pop_w()) or history (pop_h()).
class key
{
protected:
friend class mgram_map;
const int * mgram; // pointer to mgram array --key does not own that memory!
int m; // elements in mgram array
public:
// constructors
inline key() : mgram (NULL), m (0) {} // required for use in std::vector
inline key (const int * mgram, int m) : mgram (mgram), m (m) { }
// manipulations
inline key pop_h() const { if (m == 0) fail ("key::pop_h() called on empty key"); return key (mgram+1, m-1); }
inline key pop_w() const { if (m == 0) fail ("key::pop_w() called on empty key"); return key (mgram, m-1); }
// access
inline int back() const { if (m == 0) fail ("key::back() called on empty key"); return mgram[m-1]; }
inline const int & operator[] (int n) const { if (n < 0 || n >= m) fail ("key::operator[] out of bounds"); return mgram[n]; }
inline int order() const { return m; }
// key comparison (used in sorting and merging)
inline bool operator< (const key & other) const
{
for (int k = 0; k < m && k < other.m; k++)
if (mgram[k] != other.mgram[k])
return mgram[k] < other.mgram[k];
return m < other.m;
}
inline bool operator> (const key & other) const { return other < *this; }
inline bool operator<= (const key & other) const { return !(*this > other); }
inline bool operator>= (const key & other) const { return !(*this < other); }
inline bool operator== (const key & other) const
{
if (m != other.m) return false;
for (int k = 0; k < m; k++)
if (mgram[k] != other.mgram[k])
return false;
return true;
}
inline bool operator!= (const key & other) const { return !(*this == other); }
};
// 'coord' is an abstract coordinate of an m-gram. This is returned by
// operator[], and is used as an index in our sister structure, mgram_data.
struct coord
{
index_t i; // index in that level -- -1 means not found
unsigned short m; // level
inline bool valid() const { return i != nindex; }
inline void validate() const { if (!valid()) fail ("coord used but invalid"); }
void invalidate() { i = nindex; }
inline int order() const { validate(); return m; }
inline coord (int m, index_t i) : m ((unsigned short) m), i (i) {} // valid coord
// ^^ this is where we'd test for index_t overflow if we ever need it
inline coord (bool valid = true) : m (0), i (valid ? 0 : nindex) {} // root or invalid
};
// 'foundcoord' is an extended 'coord' as returned by operator[], with
// information on whether it is valid or not, and whether it refers to
// an m-gram or to a history only.
class foundcoord : public /*<-want to get rid of this*/ coord
{
const short type;
foundcoord & operator= (const foundcoord &);
public:
inline bool valid_w() const { return type > 0; }
inline bool valid_h() const { return type == 0; }
inline bool valid() const { return type >= 0; }
inline operator const coord & () const { return *this; }
inline foundcoord (short type, int m, index_t i) : type (type), coord (m, i) { }
inline foundcoord (short type) : type (type), coord (type >= 0) { }
};
// search for an mgram -- given a 'key', return its 'coord.'
// If m-gram is found, type=1. If only history found then type=0, and
// coord represents the history token instead.
// The given key may not be longer than our storage (we do not automatically
// truncate because that would not be detectable by caller).
__forceinline foundcoord operator[] (const key & k) const
{
if (k.m > M) // call truncate() first with too long keys
fail ("operator[] called with too long key");
if (k.m == 0)
return foundcoord (1); // zerogram -> root
// We traverse history one by one.
index_t i = 0;
for (int n = 1; n < k.m; n++)
{
int w = k[n -1]; // may be -1 for unknown word
int id = map (w); // may still be -1
//const char * sym = idToSymbol (id); sym; // (debugging)
i = find_child (n-1, i, id);
if (i == nindex) // unknown history: fall back
return foundcoord (-1); // indicates failure
// found it: advance search by one history token
}
// Found history. Do we also find the prediced word?
int w = k[k.m -1]; // may be -1 for unknown word
int id = map (w); // may still be -1
index_t i_m = find_child (k.m-1, i, id);
if (i_m == nindex) // not found
return foundcoord (0, k.m-1, i);
else // found
return foundcoord (1, k.m, i_m);
}
// truncate a key to the m-gram length supported by this
inline key truncate (const key & k) const { if (k.m <= M) return k; else return key (k.mgram + (k.m - M), M); }
// --- iterators
// - iterating over children of a history
// - deep-iterating over the entire tree
// for (iterator iter (mgram_map, parent_coord); iter; ++iter) { mgram_data[iter]; w=*iter; }
class iterator : public coord
{
index_t end; // end index: i is invalid when it reaches this
const mgram_map & map; // remembered for operator*
void operator=(const iterator &);
public:
// bool: true if can use or increment
inline operator bool() const { return i < end; }
// increment
inline void operator++() { if (i < end) i++; else fail ("iterator used beyond end"); }
// retrieve word -- returns -1 if not used in user's w->id map, e.g. skipped word
inline int operator*() const { if (i >= end) fail ("iterator used beyond end"); return map.id2w[map.ids[m][i]]; }
// construct 'coord' as first element
iterator (const mgram_map & map, const coord & c) : map (map)
{
c.validate();
// get the range
index_t beg = map.firsts[c.m][c.i]; // first element of child
end = map.firsts[c.m][c.i+1]; // end = first of next entry
// set the first child coordinate
m = c.m +1; // we iterate over the child level
i = beg; // first element
}
// alternative to loop over all m-grams of a level
iterator (const mgram_map & map, int m) : map (map), coord (m, 0)
{
end = (m > 0) ? (index_t) map.ids[m].size() : 1; // loop over entire vector
}
};
// for (deep_iterator iter (mgram_map, maxM); iter; ++iter) { mgram_data[iter]; key=*iter; }
class deep_iterator : public coord
{
protected:
int maxM;
std::vector<index_t> pos; // current position [0..m]
std::vector<int> mgram; // current m-gram corresponding to 'pos'
const mgram_map & map; // remembered for operator*
void operator=(const deep_iterator &);
void validate() const { if (!valid()) fail ("iterator used beyond end"); }
public:
// constructor
deep_iterator (const mgram_map & map, int p_maxM = -1)
: map (map), maxM (p_maxM), coord (map.firsts[0].size() >= 2)
{
if (maxM == -1) maxM = map.M;
else if (maxM > map.M) fail ("deep_iterator instantiated for invalid maximum depth");
mgram.resize (maxM, -1);
pos.resize (maxM + 1, 0);
}
// bool: true if can use or increment
inline operator bool() const { return valid(); }
// increment
inline void operator++()
{
validate();
// if current position has a child then enter it
if (m < maxM && m < map.M && map.firsts[m][pos[m]] < map.firsts[m][pos[m]+1])
{
i = map.firsts[m][pos[m]];
m++;
pos[m] = i;
mgram[m-1] = map.id2w[map.ids[m][i]];
return;
}
// advance vertically or step up one level
for ( ; m > 0; )
{
// advance current position if still elements left
i++;
if (i < map.firsts[m-1][pos[m-1]+1]) // not hit the end yet
{
pos[m] = i;
mgram[m-1] = map.id2w[map.ids[m][i]];
return;
}
// cannot enter or advance: step back one
m--;
i = pos[m]; // parent position
}
// reached the end
invalidate(); // invalidates 'coord'--next call to bool() will return false
return;
}
// retrieve keys -- returns -1 if not used in user's w->id map, e.g. skipped word
// The key points into the iterator structure, i.e. it operator++ invalidates it!
inline key operator*() const { validate(); return key (&mgram[0], m); }
};
// for (reordering_iterator iter (mgram_map, wrank[], maxM); iter; ++iter) { mgram_data[iter]; key=*iter; }
// Like deep_iterator, but iterates the map such that ws are returned in
// increasing wrank[w] rather than in the original storage order.
// Used for merging multiple models such as linear interpolation.
class reordering_iterator : public deep_iterator
{
const std::vector<int> & wrank; // assigns a rank to each w
const char * i; // hide coord::i against accidental access
std::vector<std::vector<index_t>> indexes; // coord::i <- indexes[m][this->i]
std::vector<index_t> indexbase; // indexes[m] is indexbase[m]-based
inline index_t & index_at (int m, index_t i)
{
return indexes[m][i - indexbase[m]];
}
std::vector<std::pair<int,int>> sortTemp; // temp for creating indexes
void operator=(const reordering_iterator &);
public:
// constructor
reordering_iterator (const mgram_map & map, const std::vector<int> & wrank, int p_maxM = -1)
: deep_iterator (map, p_maxM), wrank (wrank)
{
if (wrank.size() < map.w2id.size()) fail ("reordering_iterator: wrank has wrong dimension");
indexes.resize (maxM +1);
indexes[0].push_back (0); // look-up table for root: only one item
indexbase.resize (maxM +1, 0);
pos[0] = coord::i; // zerogram level: same i because no mapping there
if (map.M >= 1) sortTemp.reserve (map.size (1));
}
// increment
// We iterate through the map using (m, pos[m]) while user consumes (m, i)
// i.e. for operator++(), coord::i is not iterated but a return value.
inline void operator++()
{
validate();
// if current position has a child then enter it
// Note: We enter the item that coord::i points to, which is not pos[m]
// but the mapped pos[m].
if (m < maxM && m < map.M && map.firsts[m][index_at (m, pos[m])] < map.firsts[m][index_at (m, pos[m])+1])
{
// enter the level
index_t beg = map.firsts[m][index_at (m, pos[m])]; // index range of sub-level
index_t end = map.firsts[m][index_at (m, pos[m])+1];
m++;
pos[m] = beg;
// build look-up table for returned values
size_t num = end - beg;
// we sort i by rank (and i, keeping original order for identical rank)
sortTemp.resize (end - beg);
foreach_index (k, sortTemp)
{
index_t i = beg+k;
int id = map.ids[m][i];
int w = map.id2w[id];
sortTemp[k] = std::make_pair (wrank[w], i);
}
std::sort (sortTemp.begin(), sortTemp.end());
// remember sorted i's
indexbase[m] = beg; // used by index_at (m, *)
indexes[m].resize (num);
foreach_index (k, sortTemp)
index_at (m, k+beg) = sortTemp[k].second;
// set up return values
coord::i = index_at (m, pos[m]);
mgram[m-1] = map.id2w[map.ids[m][coord::i]];
return;
}
// advance vertically or step up one level
for ( ; m > 0; )
{
// advance current position if still elements left
// use our own i (in pos[m]), then map to coord::i using sorted list
pos[m]++;
if (pos[m] < map.firsts[m-1][index_at (m-1, pos[m-1])+1]) // not hit the end yet
{
coord::i = index_at (m, pos[m]);
mgram[m-1] = map.id2w[map.ids[m][coord::i]];
return;
}
// cannot enter or advance: step back one
m--;
}
// reached the end
invalidate(); // invalidates 'coord'--next call to bool() will return false
return;
}
};
// --- functions for building
// 'unmapped_key' contains original 'id' rather than 'w' values. It is only
// used for create()--at creation time, we use our private mapping.
typedef key unmapped_key;
// create a new key (to be called in sequence).
// Only the last word given in the key is added. The history of the given
// mgram must already exist and must be the last.
// Important: Unlike operator[], create() takes an unmapped_key, i.e. the
// mapping is not applied.
// 'cache' is used for speed-up, it must be as large as key.m-1 and
// initialized to 0.
#pragma warning (push) // known compiler bug: size_t (marked _w64) vs. unsigned...
#pragma warning (disable:4267) // ...int (not marked) incorrectly flagged in templates
typedef std::vector<index_t> cache_t;
coord create (const unmapped_key & k, cache_t & cache)
{
if (k.m < 1) return coord(); // (root need not be created)
// locate history (must exist), also updates cache[]
bool prevValid = true;
index_t i = 0; // index of history in level k.m-1
if (cache.empty()) cache.resize (M, nindex); // lazy initialization
for (int m = 1; m < k.m; m++)
{
int thisid = k[m-1];
if (prevValid && cache[m-1] != nindex && ids[m][cache[m-1]] == thisid)
{
i = cache[m-1]; // get from cache
continue;
}
// need to actually search
i = find_child (m-1, i, thisid);
if (i == nindex) fail ("create() called with unknown history");
cache[m-1] = i;
prevValid = false;
}
for (int m = k.m; m < M && cache[m-1] != nindex; m++)
cache[m-1] = nindex; // clear upper entries (now invalid)
// now i is the index of the id of the last history item
// make the firsts entry if not there yet
bool newHist = (firsts[k.m-1].size() < (size_t) i + 2);
while (firsts[k.m-1].size() < (size_t) i + 2) // [i+1] is the end for this array
firsts[k.m-1].push_back ((mgram_map::index_t) ids[k.m].size());
if (firsts[k.m-1].size() != (size_t) i + 2) fail ("create() called out of order (history)");
// create new word id
int thisid = k[k.m-1];
if (!newHist && thisid <= ids[k.m].back()) fail ("create() called out of order");
// keep track of idmax
if (thisid > idmax) idmax = thisid;
coord c (k.m, (index_t) ids[k.m].size());
ASSERT (firsts[k.m-1].back() == (index_t) ids[k.m].size());
ids[k.m].push_back (thisid); // create value
firsts[k.m-1].back() = (index_t) ids[k.m].size();
if (firsts[k.m-1].back() != (index_t) ids[k.m].size()) fail ("create() numeric overflow--index_t too small");
ASSERT (k.m == M || firsts[k.m].back() == (index_t) ids[k.m+1].size());
// optimization: level1nonsparse flag
// If unigram level is entirely non-sparse, we can save the search
// operation at that level, which is significantly slower than for the
// much sparser higher levels.
if (c.m == 1)
{
if (c.i == 0) level1nonsparse = true; // first entry
level1nonsparse &= (c.i == (index_t) thisid); // no search needed
level1lookup.resize (thisid +1, nindex);
level1lookup[thisid] = c.i;
}
return c;
}
#pragma warning (pop)
// call this at the end
// - establish the w->id mapping that is used in operator[]
// - finalize the firsts arrays
// This function swaps the user-provided map and our current one.
// We use swapping to avoid the memory allocation (noone else outside should
// have to keep the map).
// This function also builds our internal reverse map used in the iterator.
void created (std::vector<int> & userToLMSymMap)
{
// finalize firsts arrays
foreach_index (m, firsts)
firsts[m].resize (ids[m].size() +1, (int) ids[m+1].size());
foreach_index (m, firsts)
{
ASSERT (firsts[m][0] == 0);
foreach_index (i, ids[m])
ASSERT (firsts[m][i] <= firsts[m][i+1]);
ASSERT ((size_t) firsts[m].back() == ids[m+1].size());
}
// id mapping
// user-provided w->id map
::swap (w2id, userToLMSymMap);
// reverse map
id2w.assign (maxid()+1, nindex);
foreach_index (w, w2id)
{
int id = w2id[w];
if (id < 0) continue; // invalid word
if (id > maxid()) continue; // id not in use
id2w[id] = w;
}
}
// helper for created()--return an identical map, as we have several
// occasions where such a map is passed as userToLMSymMap to created().
std::vector<int> identical_map (size_t n = SIZE_MAX) const
{
if (n == SIZE_MAX) n = maxid() +1;
std::vector<int> v (n);
foreach_index (i, v) v[i] = i;
return v;
}
// decide whether iterator will return in increasing w order
bool inorder() const
{
#if 0 // fix this: need access to w2id, or have an inorder() function in mgram_map
bool inorder = true;
for (int i = 1; inorder && i < (int) map.w2id.size(); i++)
inorder &= (map.w2id[i+1] >= map.w2id[i]);
#endif
return false;
}
};
// ===========================================================================
// mgram_data -- data stored according to mgram_map
// Separate from mgram_map, so that we can share the same map for multiple data.
// ===========================================================================
template<class DATATYPE> class mgram_data
{
std::vector<std::vector<DATATYPE>> data;
static void fail (const char * msg) { throw runtime_error (string ("mgram_data::") + msg); }
public:
mgram_data(){}
mgram_data (int M) { init (M); }
// for an M-gram, indexes [0..M] are valid thus data[] has M+1 elements
void init (int M) { data.assign (M+1, std::vector<DATATYPE>()); }
void reserve (int m, size_t size) { data[m].reserve (size); }
void resize (int M) { if ((size_t) M+1 <= data.size()) data.resize (M+1); else fail ("resize() can only shrink"); }
size_t size (int m) const { return data[m].size(); }
size_t size() const { size_t sz = 0; foreach_index (m, data) sz += size (m); return sz; }
void clear() { data.clear(); }
void swap (mgram_data & other) { data.swap (other.data); }
// access existing elements. Usage:
// DATATYPE & element = mgram_data[mgram_map[mgram_map::key (mgram, m)]]
__forceinline DATATYPE & operator[] (const mgram_map::coord & c) { c.validate(); return data[c.m][c.i]; }
__forceinline const DATATYPE & operator[] (const mgram_map::coord & c) const { c.validate(); return data[c.m][c.i]; }
// create entire vector (for random-access situations).
void assign (int m, size_t size, const DATATYPE & value) { data[m].assign (size, value); }
// create an element. We can only append.
inline void push_back (const mgram_map::coord & c, const DATATYPE & val)
{
c.validate();
if (data[c.m].size() != (size_t) c.i) fail ("push_back() only allowed for last entry");
data[c.m].push_back (val);
}
};
// ===========================================================================
// CMGramLM -- a back-off M-gram language model in memory, loaded from an ARPA file
// ===========================================================================
class CMGramLM : public ILM
{
protected:
#if 0
void clear() // release all memory --object unusable after this
{
M = -1;
map.clear();
logP.clear();
logB.clear();
}
#endif
int M; // e.g. M=3 for trigram
// ^^ TODO: can we do away with this entirely and replace it by map.order()/this->order()
mgram_map map;
mgram_data<float> logP; // [M+1][i] probabilities
mgram_data<float> logB; // [M][i] back-off weights (stored for histories only)
friend class CMGramLMIterator;
// diagnostics of previous score() call
mutable int longestMGramFound; // longest m-gram (incl. predicted token) found
mutable int longestHistoryFound; // longest history (excl. predicted token) found
// this function is for reducing M after the fact, e.g. during estimation
// ... TODO: rethink the resize business. It is for shrinking only.
void resize (int newM)
{
M = newM;
map.resize (M);
#if 0 // ... BUGBUG: we call this before logP/logB exist
logP.resize (M);
logB.resize (M-1);
#endif
}
public:
CMGramLM() : M (-1) {} // needs explicit initialization through read() or init()
virtual int getLastLongestHistoryFound() const { return longestHistoryFound; }
virtual int getLastLongestMGramFound() const { return longestMGramFound; }
// -----------------------------------------------------------------------
// score() -- compute an m-gram score (incl. back-off and fallback)
// -----------------------------------------------------------------------
// mgram[m-1] = word to predict, tokens before that are history
// m=3 means trigram
virtual double score (const int * mgram, int m) const
{
longestHistoryFound = 0; // (diagnostics)
double totalLogB = 0.0; // accumulated back-off
for (mgram_map::key key = map.truncate (mgram_map::key (mgram, m)); ; key = key.pop_h())
{
// look up the m-gram
const mgram_map::foundcoord c = map[key];
// (diagnostics -- can be removed if not used)
if (c.valid() && key.order() -1 > longestHistoryFound)
longestHistoryFound = key.order() -1;
if (c.valid_w())
longestMGramFound = key.order();
// full m-gram found -> return it (zerogram always considered found)
if (c.valid_w())
return totalLogB + logP[c];
// history found but predicted word not -> back-off
if (c.valid_h()) // c is coordinate of parent instead
totalLogB += logB[c]; // and continue like fall back
// history not found -> fall back
} // and go again with the shortened history
}
// same as score() but without optimizations (for reference)
// ... this is really no longer needed
virtual double score_unoptimized (const int * mgram, int m) const
{ return score_unoptimized (map.truncate (mgram_map::key (mgram, m))); }
inline double score_unoptimized (const mgram_map::key & key) const
{
// look up the m-gram
const mgram_map::foundcoord c = map[key];
// full m-gram found -> return it
if (c.valid_w())
return logP[c];
// history found but predicted word not -> back-off
else if (c.valid_h()) // c is coordinate of patent instead
return logB[c] + score_unoptimized (key.pop_h());
// history not found -> fall back
else
return score_unoptimized (key.pop_h());
}
// test for OOV word (OOV w.r.t. LM)
virtual bool oov (int w) const { return map.oov (w); }
virtual void adapt (const int *, size_t) { } // this LM does not adapt
private:
// keep this for debugging
std::wstring filename; // input filename
struct SYMBOL
{
string symbol; // token
int id; // numeric id in LM space (index of word read)
bool operator< (const SYMBOL & other) const { return symbol < other.symbol; }
SYMBOL (int p_id, const char * p_symbol) : id (p_id), symbol (p_symbol) { }
};
std::vector<SYMBOL> lmSymbols; // (id, word) symbols used in LM
std::vector<int> idToSymIndex; // map LM id to index in lmSymbols[] array
// search for a word in the sorted word array.
// Only use this after sorting, i.e. after full 1-gram section has been read.
// Only really used in read().
inline int symbolToId (const char * word) const
{
int beg = 0;
int end = (int) lmSymbols.size();
while (beg < end)
{
int i = (beg + end) / 2;
const char * v = lmSymbols[i].symbol.c_str();
int cmp = strcmp (word, v);
if (cmp == 0) return lmSymbols[i].id; // found it
else if (cmp < 0) end = i; // id is left of i
else beg = i + 1; // id is right of i
}
return -1; // not found
}
inline const char * idToSymbol (int id) const
{
if (id < 0) return NULL; // empty string for unknown ids
int i = idToSymIndex[id];
return lmSymbols[i].symbol.c_str();
}
private:
// type cast to const char*, to allow write() to use both const char* and string
static const char * const_char_ptr (const char * p) { return p; }
static const char * const_char_ptr (const string & s) { return s.c_str(); }
public:
// write model out as an ARPA (text) file.
// symbols can be anything that has symbols[w] -> std::string& or const char*
template<class SYMMAP>
void write (FILE * outf, const SYMMAP & symbols,int M = INT_MAX) const
{
if (M > this->M) M = this->M; // clip; also covers default value
if (M < 1 || map.size (1) == 0)
throw runtime_error ("write: attempting to write empty model");
// output header
// \data\
// ngram 1=58289
// ngram 2=956100
// ...
fprintfOrDie (outf, "\\data\\\n");
for (int m = 1; m <= M; m++)
{
fprintfOrDie (outf, "ngram %d=%d\n", m, map.size (m));
}
fflushOrDie (outf);
// output m-grams themselves
// M-gram sections
const double log10 = log (10.0);
for (int m = 1; m <= M; m++)
{
fprintf (stderr, "estimate: writing %d %d-grams..", map.size (m), m);
int step = (int) logP.size (m) / 100;
if (step == 0) step = 1;
int numMGramsWritten = 0;
// output m-gram section
fprintfOrDie (outf, "\n\\%d-grams:\n", m);
for (mgram_map::deep_iterator iter (map, m); iter; ++iter)
{
if (iter.order() != m) // a parent
continue;
const mgram_map::key key = *iter;
ASSERT (m == key.order());
// --- output m-gram to ARPA file
fprintfOrDie (outf, "%.4f", logP[iter] / log10);
for (int k = 0; k < m; k++)
{ // the M-gram words
int wid = key[k];
const char * w = const_char_ptr (symbols[wid]);
fprintfOrDie (outf, " %s", w);
}
if (m < M)
{ // back-off weight (not for highest order)
fprintfOrDie (outf, " %.4f", logB[iter] / log10);
}
fprintfOrDie (outf, "\n");
// progress
if (numMGramsWritten % step == 0)
{
fprintf (stderr, ".");
}
numMGramsWritten++;
}
fflushOrDie (outf);
ASSERT (numMGramsWritten == map.size (m));
fprintf (stderr, "\n");
}
fprintfOrDie (outf, "\n\\end\\\n");
fflushOrDie (outf);
}
// get TopM Ngram probability
// GangLi add this function to do probability pruning
double KeepTopMNgramThreshold (int topM, int ngram)
{
// initial return as a very low value
double probThrshold = -99;
// check if nessary to prune
if (map.size(ngram) > topM)
{
std::vector<std::pair<int, float>> probArray;
probArray.reserve(map.size(ngram));
}
return probThrshold;
}
protected:
// replace zerogram prob by one appropriate for OOVs
// We use the minimum of all unigram scores (assuming they represent singleton
// events, which are closest to a zerogram--a better choice may be a leaving-
// one-out estimate?).
// Back-off weight is reset to 1.0 such that there is no extra penalty on it.
void updateOOVScore()
{
float unknownLogP = 0.0f;
for (mgram_map::iterator iter (map, mgram_map::coord()); iter; ++iter)
{
if (logP[iter] < -98.0f) continue; // disabled token, such as <s>, does not count
if (logP[iter] < unknownLogP)
unknownLogP = logP[iter];
}
logP[mgram_map::coord()] = unknownLogP;
logB[mgram_map::coord()] = 0.0f;
}
public:
// read an ARPA (text) file.
// Words do not need to be sorted in the unigram section, but the m-gram
// sections have to be in the same order as the unigrams.
// The 'userSymMap' defines the vocabulary space used in score().
// If 'filterVocabulary' then LM entries for words not in userSymMap are skipped.
// Otherwise the userSymMap is updated with the words from the LM.
// 'maxM' allows to restrict the loading to a smaller LM order.
// SYMMAP can be e.g. CSymMap or CSymbolSet.
template<class SYMMAP>
void read (const std::wstring & pathname, SYMMAP & userSymMap, bool filterVocabulary, int maxM)
{
int lineNo = 0;
msra::basetypes::auto_file_ptr f = fopenOrDie (pathname, L"rbS");
fprintf (stderr, "read: reading %S", pathname.c_str());
filename = pathname; // (keep this info for debugging)
// --- read header information
// search for header line
char buf[1024];
lineNo++, fgetline (f, buf);
while (strcmp (buf, "\\data\\") != 0 && !feof (f))
lineNo++, fgetline (f, buf);
lineNo++, fgetline (f, buf);
// get the dimensions
std::vector<int> dims; dims.reserve (4);
while (buf[0] == 0 && !feof (f))
lineNo++, fgetline (f, buf);
int n, dim;
dims.push_back (1); // dummy zerogram entry
while (sscanf (buf, "ngram %d=%d", &n, &dim) == 2 && n == (int) dims.size())
{
dims.push_back (dim);
lineNo++, fgetline (f, buf);
}
M = (int) dims.size() -1;
if (M == 0)
RuntimeError ("read: mal-formed LM file, no dimension information (%d): %S", lineNo, pathname.c_str());
int fileM = M;
if (M > maxM)
M = maxM;
// allocate main storage
map.init (M);
logP.init (M);
logB.init (M-1);
for (int m = 0; m <= M; m++)
{
map.reserve (m, dims[m]);
logP.reserve (m, dims[m]);
if (m < M)
logB.reserve (m, dims[m]);
}
lmSymbols.reserve (dims[0]);
logB.push_back (mgram_map::coord(), 0.0f); // dummy logB for backing off to zg
logP.push_back (mgram_map::coord(), 0.0f); // zerogram score -- gets updated later
std::vector<bool> skipWord; // true: skip entry containing this word
skipWord.reserve (lmSymbols.capacity());
// --- read main sections
const double ln10xLMF = log (10.0); // ARPA scores are strangely scaled
msra::strfun::tokenizer tokens (" \t\n\r", M+1); // used in tokenizing the input line
for (int m = 1; m <= M; m++)
{
while (buf[0] == 0 && !feof (f))
lineNo++, fgetline (f, buf);
if (sscanf (buf, "\\%d-grams:", &n) != 1 || n != m)
RuntimeError ("read: mal-formed LM file, bad section header (%d): %S", lineNo, pathname.c_str());
lineNo++, fgetline (f, buf);
std::vector<int> mgram (m +1, -1); // current mgram being read ([0]=dummy)
std::vector<int> prevmgram (m +1, -1); // cache to speed up symbol lookup
mgram_map::cache_t mapCache; // cache to speed up map.create()
// read all the m-grams
while (buf[0] != '\\' && !feof (f))
{
if (buf[0] == 0)
{
lineNo++, fgetline (f, buf);
continue;
}
// -- parse the line
tokens = &buf[0];
if ((int) tokens.size() != ((m < fileM) ? m + 2 : m + 1))
RuntimeError ("read: mal-formed LM file, incorrect number of tokens (%d): %S", lineNo, pathname.c_str());
double scoreVal = atof (tokens[0]); // ... use sscanf() instead for error checking?
double thisLogP = scoreVal * ln10xLMF; // convert to natural log
bool skipEntry = false;
for (int n = 1; n <= m; n++)
{
const char * tok = tokens[n];
// map to id
int id;
if (m == 1) // unigram: build vocab table
{
id = (int) lmSymbols.size(); // unique id for this symbol
lmSymbols.push_back (SYMBOL (id, tok));
bool toSkip = false;
if (userSymMap.sym2existingId (lmSymbols.back().symbol) == -1)
{
if (filterVocabulary)
toSkip = true; // unknown word
else
userSymMap.sym2id (lmSymbols.back().symbol); // create it in user's space
}
skipWord.push_back (toSkip);
}
else // mgram: look up word in vocabulary
{
if (prevmgram[n] >= 0 && strcmp (idToSymbol (prevmgram[n]), tok) == 0)
id = prevmgram[n]; // optimization: most of the time, it's the same
else
{
id = symbolToId (tok);
if (id == -1)
RuntimeError ("read: mal-formed LM file, m-gram contains unknown word (%d): %S", lineNo, pathname.c_str());
}
}
mgram[n] = id; // that's our id
skipEntry |= skipWord[id]; // skip entry if any token is unknown
}
double thisLogB = 0.0;
if (m < M && !skipEntry)
{
double boVal = atof (tokens[m+1]); // ... use sscanf() instead for error checking?
thisLogB = boVal * ln10xLMF; // convert to natural log
}
lineNo++, fgetline (f, buf);
if (skipEntry) // word contained unknown vocabulary: skip entire entry
goto skipMGram;
// -- enter the information into our data structure
// Note that the mgram_map/mgram_data functions are highly efficient
// because they can only be called in sorted order.
// locate the corresponding entries
{ // (local block because we 'goto' over this)
mgram_map::key key (&mgram[1], m); // key to locate this m-gram
mgram_map::coord c = map.create (key, mapCache);// create it & gets its location
// enter into data structure
logP.push_back (c, (float) thisLogP); // prob value
if (m < M) // back-off weight
logB.push_back (c, (float) thisLogB);
}
skipMGram:
// remember current mgram for next iteration
::swap (mgram, prevmgram);
}
// fix the symbol set -- now we can binary-search in them with symbolToId()
if (m == 1)
{
std::sort (lmSymbols.begin(), lmSymbols.end());
idToSymIndex.resize (lmSymbols.size(), -1);
for (int i = 0; i < (int) lmSymbols.size(); i++)
{
idToSymIndex[lmSymbols[i].id] = i;
}
}
fprintf (stderr, ", %d %d-grams", map.size (m), m);
}
fprintf (stderr, "\n");
// check end tag
if (M == fileM)
{ // only if caller did not restrict us to a lower order
while (buf[0] == 0 && !feof (f))
lineNo++, fgetline (f, buf);
if (strcmp (buf, "\\end\\") != 0)
RuntimeError ("read: mal-formed LM file, no \\end\\ tag (%d): %S", lineNo, pathname.c_str());
}
// update zerogram score by one appropriate for OOVs
updateOOVScore();
// establish mapping of word ids from user to LM space.
// map's operator[] maps mgrams using this map.
std::vector<int> userToLMSymMap (userSymMap.size());
for (int i = 0; i < (int) userSymMap.size(); i++)
{
const char * sym = userSymMap.id2sym (i);
int id = symbolToId (sym); // may be -1 if not found
userToLMSymMap[i] = id;
}
map.created (userToLMSymMap);
}
protected:
// sort LM such that iterators will iterate in increasing order w.r.t. w2id[w]
// This is achieved by replacing all internal ids by w2id[w].
// This function is expensive: it makes a full temporary copy and involves sorting.
// w2id[] gets destroyed by this function.
void sort (std::vector<int> & w2id)
{
// create a full copy of logP and logB in the changed order
mgram_map sortedMap (M);
mgram_data<float> sortedLogP (M);
mgram_data<float> sortedLogB (M-1);
for (int m = 1; m <= M; m++)
{
sortedMap.reserve (m, map.size (m));
sortedLogP.reserve (m, logP.size (m));
if (m < M) sortedLogB.reserve (m, logB.size (m));
}
// iterate in order of w2id
// Order is determined by w2id[], i.e. entries with lower new id are
// returned first.
std::vector<int> mgram (M+1, -1); // unmapped key in new id space
mgram_map::cache_t createCache;
for (mgram_map::reordering_iterator iter (map, w2id); iter; ++iter)
{
int m = iter.order();
mgram_map::key key = *iter; // key in old 'w' space
// keep track of an unmapped key in new id space
if (m > 0)
{
int w = key.back();
int newid = w2id[w]; // map to new id space
mgram[m-1] = newid;
}
for (int k = 0; k < m; k++) ASSERT (mgram[k] == w2id[key[k]]);
// insert new key into sortedMap
mgram_map::coord c = sortedMap.create (mgram_map::unmapped_key (&mgram[0], m), createCache);
// copy over logP and logB
sortedLogP.push_back (c, logP[iter]);
if (m < M)
sortedLogB.push_back (c, logB[iter]);
}
// finalize sorted map
sortedMap.created (w2id);
// replace LM by sorted LM
map.swap (sortedMap);
logP.swap (sortedLogP);
logB.swap (sortedLogB);
}
public:
// sort LM such that internal ids are in lexical order
// After calling this function, iterators will iterate in lexical order,
// and writing to an ARPA file creates a lexicographically sorted file.
// Having sorted files is useful w.r.t. efficiency when iterating multiple
// models in parallel, e.g. interpolating or otherwise merging models,
// because then IIter can use the efficient deep_iterator (which iterates
// in our internal order and therefore does not do any sorting) rather than
// the reordering_iterator (which involves sort operations).
template<class SYMMAP>
void sort (const SYMMAP & userSymMap)
{
// deterine sort order
// Note: This code copies all strings twice.
std::vector<pair<std::string, int>> sortTemp (userSymMap.size()); // (string, w)
foreach_index (w, sortTemp)
sortTemp[w] = make_pair (userSymMap[w], w);
std::sort (sortTemp.begin(), sortTemp.end());
std::vector<int> w2id (userSymMap.size(), -1); // w -> its new id
foreach_index (id, w2id)
w2id[sortTemp[id].second] = id;
// sort w.r.t. new id space
sort (w2id);
}
// iterator to enumerate all known m-grams
// This is used when creating whole models at once.
template<class ITERATOR>
class TIter : public ILM::IIter
{
int minM; // minimum M we want to iterate (skip all below)
const CMGramLM & lm; // the underlying LM (for value())
std::vector<int> wrank; // sorting criterion
ITERATOR iter; // the iterator used in this interface
void findMinM() { while (iter && iter.order() < minM) ++iter; }
public:
// constructors
TIter (const CMGramLM & lm, int minM, int maxM)
: minM (minM), lm (lm), iter (lm.map, maxM)
{ findMinM(); }
TIter (const CMGramLM & lm, bool, int minM, int maxM)
: minM (minM), lm (lm), wrank (lm.map.identical_map (lm.map.maxw()+1)),
iter (lm.map, wrank, maxM)
{ findMinM(); }
// has iterator not yet reached end?
virtual operator bool() const { return iter; }
// advance by one
virtual void operator++()
{
++iter;
findMinM();
}
// current m-gram (mgram,m)
virtual std::pair<const int*,int> operator*() const
{
mgram_map::key key = *iter;
return std::make_pair (key.order() == 0 ? NULL : &key[0], key.order());
}
// current value (logP, logB)
// No processing here--read out the logP/logB values directly from the data structure.
virtual std::pair<double,double> value() const
{
if (iter.order() < lm.M)
return std::make_pair (lm.logP[iter], lm.logB[iter]);
else
return std::make_pair (lm.logP[iter], 0.0);
}
};
virtual IIter * iter (int minM, int maxM) const
{
if (maxM == INT_MAX) maxM = M; // default value
// if no sorting needed, then we can use the efficient deep_iterator
if (map.inorder())
return new TIter<mgram_map::deep_iterator> (*this, minM, maxM);
// sorting needed: use reordering_iterator
return new TIter<mgram_map::reordering_iterator> (*this, true, minM, maxM);
}
virtual int order() const { return M; }
virtual size_t size (int m) const { return (int) logP.size (m); }
protected:
// computeSeenSums -- compute sum of seen m-grams, store at their history coord
// If islog then P is logP, otherwise linear (non-log) P.
template<class FLOATTYPE>
static void computeSeenSums (const mgram_map & map, int M, const mgram_data<float> & P,
mgram_data<FLOATTYPE> & PSum, mgram_data<FLOATTYPE> & backoffPSum,
bool islog)
{
// dimension the accumulators and initialize them to 0
PSum.init (M-1);
for (int m = 0; m <= M-1; m++) PSum.assign (m, map.size (m), 0);
backoffPSum.init (M-1);
for (int m = 0; m <= M-1; m++) backoffPSum.assign (m, map.size (m), 0);
// iterate over all seen m-grams
msra::basetypes::fixed_vector<mgram_map::coord> histCoord (M); // index of history mgram
for (mgram_map::deep_iterator iter (map, M); iter; ++iter)
{
int m = iter.order();
if (m < M) histCoord[m] = iter;
if (m == 0) continue;
const mgram_map::key key = *iter;
ASSERT (m == key.order());
float thisP = P[iter];
if (islog)
{
if (thisP <= logzero) continue; // pruned or otherwise lost
thisP = exp (thisP);
}
else
{
if (thisP == 0.0f) continue; // a pruned or otherwise lost m-gram
}
// parent entry
const mgram_map::coord j = histCoord[m-1]; // index of parent entry
// accumulate prob in B field (temporarily misused)
PSum[j] += thisP;
// the mass of the back-off distribution covered by higher-order seen m-grams.
// This must exist, as any sub-sequence of any seen m-mgram exists
// due to the way we count the tokens.
const mgram_map::key boKey = key.pop_h();
const mgram_map::foundcoord c = map[boKey];
if (!c.valid_w())
throw runtime_error ("estimate: malformed data: back-off value not found"); // must exist
// look it up
float Pc = P[c];
backoffPSum[j] += islog ? exp (Pc) : Pc;
}
}
// computeBackoff -- compute back-off weights
// Set up or update logB[] based on P[].
// logB[] is an output from this function only.
// If islog then P is logP, otherwise linear (non-log) P.
static void computeBackoff (const mgram_map & map, int M,
const mgram_data<float> & P, mgram_data<float> & logB,
bool islog)
{
mgram_data<float> backoffPSum; // accumulator for the probability mass covered by seen m-grams
// sum up probabilities of seen m-grams
// - we temporarily use the B field for the actual seen probs
// - and backoffSum for their prob pretending we are backing off
computeSeenSums (map, M, P, logB, backoffPSum, islog);
// That has dimensioned logB as we need it.
// derive the back-off weight from it
for (mgram_map::deep_iterator iter (map, M-1); iter; ++iter)
{
double seenMass = logB[iter]; // B field misused: sum over all seen children
if (seenMass > 1.0)
{
if (seenMass > 1.0001) // (a minor round-off error is acceptable)
fprintf (stderr, "estimate: seen mass > 1.0: %8.5f --oops??\n", seenMass);
seenMass = 1.0; // oops?
}
// mass covered by seen m-grams is unused -> take out
double coveredBackoffMass = backoffPSum[iter];
if (coveredBackoffMass > 1.0)
{
if (coveredBackoffMass > 1.0001) // 1.0 for unigrams, sometimes flags this
fprintf (stderr, "estimate: unseen backoff mass < 0: %8.5f --oops??\n", 1.0 - coveredBackoffMass);
coveredBackoffMass = 1.0; // oops?
}
// redistribute such that
// seenMass + bow * usedBackoffMass = 1
// ==> bow = (1 - seenMass) / usedBackoffMass
double freeMass = 1.0 - seenMass;
double accessibleBackoffMass = 1.0 - coveredBackoffMass; // sum of all backed-off items
// back-off weight is just the free probability mass
double bow = (accessibleBackoffMass > 0) ? freeMass / accessibleBackoffMass : 1.0;
// A note on the curious choice of bow=1.0 for accessibleBackoffMass==0:
// If accessibleBackoffMass==0, we are in undefined territory.
// Because this means we never back off. Problem is that we have
// already discounted the probabilities, i.e. there is probability
// mass missing (distribution not normalized). Possibilities for
// remedying the normalization issue are:
// 1. use linear interpolation instead generally
// 2. use linear interpolation only for such distributions
// 3. push mass into <UNK> class if available
// 4. ignore the normalization problem.
// We choose 2. for the unigram distribution (enforced outside of this
// function), and 4. for all other cases.
// A second question arises for OOV words in this case. With OOVs,
// accessibleBackoffMass should no longer be 0, but we don't know its
// value. Be Poov the mass of all OOV words, then
// bow = (1 - seenMass) / Poov
// Further, if seenMass was not discounted (as in our unigram case),
// it computes to 1, but if we had accounted for Poov, it would
// compute as (1-Poov) instead. Thus,
// bow = (1 - (1-Poov)) / Poov = 1
// Realistically, this case happens for the unigram distribution.
// Practically it means fallback instead of back-off for OOV words.
// Also, practically, Poov is very small, so is the error.
logB[iter] = logclip ((float) bow);
}
}
};
// ===========================================================================
// CMGramLMIterator -- a special-purpose class that allows for direct iteration.
// ===========================================================================
class CMGramLMIterator : public msra::lm::mgram_map::iterator
{
const CMGramLM & lm;
public:
CMGramLMIterator (const CMGramLM & lm, mgram_map::coord c) : lm (lm), msra::lm::mgram_map::iterator (lm.map, c) {}
float logP() const { return lm.logP[*this]; }
float logB() const { return lm.logB[*this]; }
float logB (mgram_map::coord c) const { return lm.logB[c]; }
msra::lm::mgram_map::coord locate (const int * mgram, int m) const
{
msra::lm::mgram_map::foundcoord c = lm.map[msra::lm::mgram_map::key (mgram, m)];
if (!c.valid_w())
throw std::logic_error ("locate: attempting to locate a non-existing history");
return c;
}
};
// ===========================================================================
// CMGramLMEstimator -- estimator for CMGramLM
// Implements Kneser-Ney discounting with Goodman/Chen modification, as well
// as Kneser-Ney back-off.
// ===========================================================================
class CMGramLMEstimator : public CMGramLM
{
mgram_data<unsigned int> counts; // [M+1][i] counts
mgram_map::cache_t mapCache; // used in map.create()
std::vector<int> adaptBuffer; // adapt() pushes data in here
std::vector<int> adaptBufferHead; // first M-1 tokens for cyclic closure
std::vector<unsigned int> minObs; // GangLi: prune each gram by obs occur
public:
// calling sequence:
// - init()
// - push_back() for each count, in increasing order
// - estimate() -- heavy lifting happens here
// - writeARPA() to file
// ... missing: forms of score-based pruning should happen here
// construct
void init (int p_M)
{
// dimensions
M = p_M;
map.init (M);
logP.clear();
logB.clear();
counts.init (M);
if ((int) minObs.size() != M)
{// first time initial
minObs.resize(M, 0);
if (M > 2) minObs[2] = 2; // GangLi: prune trigram if Obs < 2, this is default value
fprintf (stderr, "Set miniObs to 0 0 2.\n");
}
else
{
fprintf (stderr, "Not reset miniObs because it has already been set.\n");
}
for (int m = 1; m <= M; m++) counts.reserve (m, 1000000); // something to start with
}
// set std::vector<unsigned int> minObs
void setMinObs(const std::vector<unsigned int> & setMinObs)
{
if (minObs.size() != setMinObs.size())
RuntimeError("In setMinObs: setMinObs size (%d) is not for %d-gram.", setMinObs.size(), minObs.size());
minObs = setMinObs;
}
// call count() repeatedly to add counts, then call estimate() when done.
// Include history counts. Probabilities are based on provided history
// counts, rather than the sum of seen m-grams, to allow for count pruning.
void push_back (const int * mgram, int m, unsigned int count)
{
if (m > M) throw runtime_error ("push_back: called with m-gram longer than M");
// add to mgram_map & get location
mgram_map::coord c = map.create (mgram_map::unmapped_key (mgram, m), mapCache);
// save the count
counts.push_back (c, count);
};
protected:
// add all tokens from adaptBuffer to counts[].
// This is an expensive operation involving recreating the map, so use
// this only for large chunks of data at a time.
void merge()
{
// we need at least one M-gram
int ntoks = (int) adaptBuffer.size() - (M -1);
if (ntoks < 1)
return;
// create sorted set of counts and create merged counts
mgram_map mmap (M);
mgram_data<unsigned int> mcounts (M);
mcounts.push_back (mgram_map::coord(), ntoks); // zerogram count
std::vector<int> keybuf (M+1);
// do one order after another (to save memory)
fprintf (stderr, "merge: adding %d tokens...", ntoks);
for (int m = 1; m <= M; m++)
{
mgram_map::cache_t mmapCache;
// enumerate all m-grams of this order
std::vector<mgram_map::key> keys (ntoks);
foreach_index (j, keys) keys[j] = mgram_map::key (&adaptBuffer[j], m);
// sort them
std::sort (keys.begin(), keys.end());
// pre-allocate
size_t alloc = counts.size (m);
alloc++; // count first key
for (int j = 1; j < ntoks; j++)
if (keys[j] > keys[j-1])
alloc++; // count unique keys
mmap.reserve (m, alloc); // worst case: no overlap
mcounts.reserve (m, alloc);
// merge with existing counts
// Typical merge-sort operation with two iterators.
mgram_map::deep_iterator iter (map, m);
int i = 0;
while (i < ntoks || iter)
{
if (iter && iter.m != m)
{
++iter;
continue;
}
if (iter)
{
// regular case (neither has reached the end)
if (i < ntoks && iter)
{
// Note: *iter is a 'mapped' key, while create() expects an
// unmapped one. During training, both are identical.
mgram_map::unmapped_key oldkey = (mgram_map::unmapped_key) *iter;
if (oldkey < keys[i]) // key exists in old counts but not new
{
unsigned int count = counts[iter];
mcounts.push_back (mmap.create (oldkey, mmapCache), count);// store 'count' under 'key'
++iter; // watch out: this invalidates oldkey
}
else
{
// a range of new m-grams
mgram_map::unmapped_key newkey = keys[i];
unsigned int count = 1;
i++;
while (i < ntoks && newkey == keys[i])
{ // consume new tokens with the same key
count++;
i++;
}
if (oldkey == newkey) // if old mgram matches then consume it
{
count += counts[iter]; // sum both up
++iter;
}
mcounts.push_back (mmap.create (newkey, mmapCache), count);
}
}
else // if (i == ntoks && iter)
{ // final old counts
unsigned int count = counts[iter];
mgram_map::unmapped_key oldkey = (mgram_map::unmapped_key) *iter;
mcounts.push_back (mmap.create (oldkey, mmapCache), count);
++iter;
}
}
else // if (i < ntoks && !iter)
{ // final new counts
mgram_map::unmapped_key newkey = keys[i];
unsigned int count = 1;
i++;
while (i < ntoks && newkey == keys[i])
{ // consume new tokens with the same key
count++;
i++;
}
mcounts.push_back (mmap.create (newkey, mmapCache), count); // store 'count' under 'key'
}
}
fprintf (stderr, " %d %d-grams", mcounts.size (m), m);
}
// remove used up tokens from the buffer
adaptBuffer.erase (adaptBuffer.begin(), adaptBuffer.begin() + ntoks);
// Establish w->id mapping -- mapping is identical (w=id) during estimation.
std::vector<int> w2id (mmap.maxid() +1);
foreach_index (i, w2id) w2id[i] = i;
//std::vector<int> w2id (mmap.identical_map());
// close down creation of new tokens, so we can random-access
mmap.created (w2id);
// and swap
map.swap (mmap);
counts.swap (mcounts);
fprintf (stderr, "\n");
// destructor will delete previous counts and map (now in mcount/mmap)
}
public:
// training by pushing data in
// special modes:
// - data=NULL -> reset; m=LM order from now on
// - data not NULL but m=0 -> taken as end indicator
virtual void adapt (const int * data, size_t m)
{
// special call for reset
if (data == NULL)
{
if (m == 0) throw runtime_error ("adapt: must pass LM order");
init ((int) m); // clear out current LM
adaptBuffer.clear();
adaptBufferHead.clear();
return;
}
// special call to flush (incl. cyclic closure)
if (m == 0)
{
// cyclicaly close the data set if final
adaptBuffer.insert (adaptBuffer.end(), adaptBufferHead.begin(), adaptBufferHead.end());
adaptBufferHead.clear();
// merge the remaining tokens in
merge();
adaptBuffer.clear(); // the cyclically closed tokens remain->clear
return;
}
// regular call: pushing word tokens in
const size_t countChunkSize = 10000000; // 10 million
adaptBuffer.reserve (countChunkSize);
// insert into our buffer
adaptBuffer.insert (adaptBuffer.end(), data, data + m);
// remember initial tokens for cyclic closure
while (m > 0 && (int) adaptBufferHead.size() < M-1)
{
adaptBufferHead.push_back (*data);
data++;
m--;
}
// flush the buffer
if (adaptBuffer.size() > countChunkSize)
merge();
}
#if 0 // debugging code -- rename adapt() above to adapt1()
virtual void adapt (const int * data, size_t m)
{
while (m > 2)
{
adapt1 (data, 2);
data += 2;
m -= 2;
}
while (m > 0)
{
adapt1 (data, 1);
data++;
m--;
}
}
#endif
protected:
// read one file
// If dropId != -1 then do not create userSymMap but look up entries, and
// use dropId for all unknown ones.
template<class SYMMAP>
int read (FILE * f, SYMMAP & userSymMap, int startId, int endId, int dropId)
{
const SYMMAP & constSymMap = userSymMap;
// fgetline will check the line length, so enlarge the buf.size
std::vector<char> buf (5000000);
std::vector<int> ids;
ids.reserve (buf.size() / 4);
msra::strfun::tokenizer tokens (" \t", ids.capacity());
int totalTokens = 0; // for visual feedback
while (!feof (f))
{
// fgetline will check the line length, so enlarge the buf.size
tokens = fgetline (f, &buf[0], (int) buf.size());
if (tokens.empty()) continue;
ids.resize (0);
ids.push_back (startId);
foreach_index (i, tokens)
{
const char * p = tokens[i];
int id = dropId == -1 ? userSymMap[p] : constSymMap[p];
ids.push_back (id);
if (totalTokens++ % 100000 == 0) fprintf (stderr, ".");
}
ids.push_back (endId);
totalTokens += 2;
adapt (&ids[0], ids.size());
}
return totalTokens;
}
public:
// 'read' here means read text.
// filterVocabulary:
// false - no filter. The userSymMapis built (or augmented) by this function,
// incl. sentence boundary markers <s> and </s>
// true - remove all words that are not in userSymMap. The userSymMap is
// not modified. If <UNK> is present, unknown words are mapped to
// it. Otherwise, m-grams involving OOV words are pruned.
template<class SYMMAP>
void read (const std::wstring & pathname, SYMMAP & userSymMap, bool filterVocabulary, int maxM)
{
if (!filterVocabulary)
{ // create <s> and </s>
userSymMap["<s>"];
userSymMap["</s>"];
}
const SYMMAP & constSymMap = userSymMap;
int startId = constSymMap["<s>"]; // or -1 -- but later enforce it is not
int endId = constSymMap["</s>"]; // or -1 -- dito.
int unkId = constSymMap["<UNK>"]; // or -1 -- -1 is OK
if (startId == -1 || endId == -1) // if filtering, these must be given
throw runtime_error ("read: <s> and/or </s> missing in vocabulary");
// if filtering but no <UNK>, we use (vocabsize) as the id, and have
// estimate() prune it
int dropId = filterVocabulary ? unkId != -1 ? unkId : userSymMap.size() : -1;
if (filterVocabulary)
RuntimeError ("CMGramLMEstimator::read() not tested for filterVocabulary==true");
// reset adaptation
adapt (NULL, maxM); // pass dimension here
// read all material and push into counts
msra::basetypes::auto_file_ptr f = fopenOrDie (pathname, L"rbS");
std::string tag = fgetline (f);
if (tag == "#traintext")
{
read (f, userSymMap, startId, endId, dropId);
}
else if (tag == "#trainfiles")
{
while (!feof (f))
{
string thispath = fgetline (f);
if (thispath.empty() || thispath[0] == '#') continue; // comment
msra::basetypes::auto_file_ptr thisf = fopenOrDie (thispath, "rbS");
fprintf (stderr, "read: ingesting training text from %s ..", thispath.c_str());
int numTokens = read (thisf, userSymMap, startId, endId, dropId);
fprintf (stderr, "%d tokens\n", numTokens);
}
}
else if (!tag.empty() && tag[0] == '#')
{
RuntimeError ("read: unknown tag '%s'", tag.c_str());
}
else // no tag: just load the file directly
{
rewind (f);
read (f, userSymMap, startId, endId, dropId);
}
// finalize
adapt (&maxM, 0);
// estimate
vector<bool> dropWord (userSymMap.size(), false);
dropWord.push_back (true); // filtering but no <UNK>:
ASSERT (!filterVocabulary || unkId != -1 || dropWord[dropId]);
//std::vector<unsigned int> minObs (2, 0);
//std::vector<unsigned int> iMinObs (3, 0);
//iMinObs[1] = 3; // remove singleton 2+-grams
//iMinObs[2] = 3; // remove singleton 3+-grams
//// set prune value to 0 3 3
//setMinObs (iMinObs);
for (size_t i = 0; i < minObs.size(); i++)
{
MESSAGE("minObs %d: %d.", i, minObs[i]);
}
estimate (startId, minObs, dropWord);
#if 0 // write it out for debugging
vector<string> syms (userSymMap.size());
foreach_index (i, syms) syms[i] = userSymMap[i];
auto_file_ptr outf = fopenOrDie ("d:/debug.lm", "wbS");
write (outf, syms);
#endif
}
protected:
// reduce M
void resize (int newM)
{
CMGramLM::resize (newM);
counts.resize (newM);
}
public:
// -----------------------------------------------------------------------
// estimate() -- estimate a back-off m-gram language model.
// -----------------------------------------------------------------------
// - Kneser-Ney absolute discounting
// - Goodman-Shen count-specific discounting values
// - Kneser-Ney back-off
// minObs is 0-based, i.e. minObs[0] is the cut-off for unigrams.
void estimate (int startId, const std::vector<unsigned int> & minObs, vector<bool> dropWord)
{
if (!adaptBuffer.empty())
throw runtime_error ("estimate: adaptation buffer not empty, call adapt(*,0) to flush buffer first");
// Establish w->id mapping -- mapping is identical (w=id) during estimation.
std::vector<int> w2id (map.maxid() +1);
foreach_index (i, w2id) w2id[i] = i;
//std::vector<int> w2id (map.identical_map());
// close down creation of new tokens, so we can random-access
map.created (w2id);
// ensure M reflects the actual order of read data
while (M > 0 && counts.size (M) == 0) resize (M-1);
for (int m = 1; m <= M; m++)
fprintf (stderr, "estimate: read %d %d-grams\n", counts.size (m), m);
// === Kneser-Ney smoothing
// This is a strange algorithm.
#if 1 // Kneser-Ney back-off
// It seems not to work for fourgram models (applied to the trigram).
// But if it is only applied to bigram and unigram, there is a gain
// from the fourgram. So we are not applying it to trigram and above.
// ... TODO: use a constant to define the maximum KN count level,
// and then do not allocate memory above that.
mgram_data<unsigned int> KNCounts; // [shifted m-gram] (*,v,w)
mgram_data<unsigned int> KNTotalCounts; // [shifted, shortened m-gram] (*,v,*)
if (M >= 2)
{
fprintf (stderr, "estimate: allocating Kneser-Ney counts...\n");
KNCounts.init (M-1);
for (int m = 0; m <= M-1; m++) KNCounts.assign (m, counts.size (m), 0);
KNTotalCounts.init (M-2);
for (int m = 0; m <= M-2; m++) KNTotalCounts.assign (m, counts.size (m), 0);
fprintf (stderr, "estimate: computing Kneser-Ney counts...\n");
// loop over all m-grams to determine KN counts
for (mgram_map::deep_iterator iter (map); iter; ++iter)
{
const mgram_map::key key = *iter;
if (key.order() < 2) continue; // undefined for unigrams
const mgram_map::key key_w = key.pop_h();
const mgram_map::foundcoord c_w = map[key_w];
if (!c_w.valid_w())
throw runtime_error ("estimate: invalid shortened KN m-gram");
KNCounts[c_w]++; // (u,v,w) -> count (*,v,w)
const mgram_map::key key_h = key_w.pop_w();
mgram_map::foundcoord c_h = map[key_h];
if (!c_h.valid_w())
throw runtime_error ("estimate: invalid shortened KN history");
KNTotalCounts[c_h]++; // (u,v,w) -> count (*,v,w)
}
}
#else // regular back-off: just use regular counts instad
mgram_data<unsigned int> & KNCounts = counts;
mgram_data<unsigned int> & KNTotalCounts = counts;
// not 'const' so we can later clear() them... this is only for testng anyway
#endif
// === estimate "modified Kneser-Ney" discounting values
// after Chen and Goodman: An empirical study of smoothing techniques for
// language modeling, CUED TR-09-09 -- a rich resource about everything LM!
std::vector<double> d1 (M+1, 0.0);
std::vector<double> d2 (M+1, 0.0);
std::vector<double> d3 (M+1, 0.0);
fprintf (stderr, "estimate: discounting values:");
{
// actually estimate discounting values
std::vector<int> n1 (M+1, 0); // how many have count=1, 2, 3, 4
std::vector<int> n2 (M+1, 0);
std::vector<int> n3 (M+1, 0);
std::vector<int> n4 (M+1, 0);
for (mgram_map::deep_iterator iter (map); iter; ++iter)
{
int m = iter.order();
if (m == 0) continue; // skip the zerogram
unsigned int count = counts[iter];
// Kneser-Ney smoothing can also be done for back-off weight computation
if (m < M && m < 3) // for comments see where we estimate the discounted probabilities
{ // ^^ seems not to work for 4-grams...
const mgram_map::key key = *iter; // needed to check for startId
ASSERT (key.order() == m);
if (m < 2 || key.pop_w().back() != startId)
{
count = KNCounts[iter];
if (count == 0) // must exist
throw runtime_error ("estimate: malformed data: back-off value not found (numerator)");
}
}
if (count == 1) n1[m]++;
else if (count == 2) n2[m]++;
else if (count == 3) n3[m]++;
else if (count == 4) n4[m]++;
}
for (int m = 1; m <= M; m++)
{
if (n1[m] == 0) throw runtime_error (msra::strfun::strprintf ("estimate: error estimating discounting values: n1[%d] == 0", m));
if (n2[m] == 0) throw runtime_error (msra::strfun::strprintf ("estimate: error estimating discounting values: n2[%d] == 0", m));
//if (n3[m] == 0) RuntimeError ("estimate: error estimating discounting values: n3[%d] == 0", m);
double Y = n1[m] / (n1[m] + 2.0 * n2[m]);
if (n3[m] ==0 || n4[m] == 0)
{
fprintf (stderr, "estimate: n3[%d] or n4[%d] is 0, falling back to unmodified discounting\n", m, m);
d1[m] = Y;
d2[m] = Y;
d3[m] = Y;
}
else
{
d1[m] = 1.0 - 2.0 * Y * n2[m] / n1[m];
d2[m] = 2.0 - 3.0 * Y * n3[m] / n2[m];
d3[m] = 3.0 - 4.0 * Y * n4[m] / n3[m];
}
// ... can these be negative??
fprintf (stderr, " (%.3f, %.3f, %.3f)", d1[m], d2[m], d3[m]);
}
fprintf (stderr, "\n");
}
// === threshold against minimum counts (set counts to 0)
// this is done to save memory, but it has no impact on the seen probabilities
// ...well, it does, as pruned mass get pushed to back-off distribution... ugh!
fprintf (stderr, "estimate: pruning against minimum counts...\n");
// prune unigrams first (unigram cut-off can be higher than m-gram cut-offs,
// as a means to decimate the vocabulary)
unsigned int minUniObs = minObs[0]; // minimum unigram count
int removedWords = 0;
for (mgram_map::iterator iter (map, 1); iter; ++iter)
{ // unigram pruning is special: may be higher than m-gram threshold
if (counts[iter] >= minUniObs) continue;
int wid = *iter;
dropWord[wid] = true; // will throw out all related m-grams
removedWords++;
}
fprintf (stderr, "estimate: removing %d too rare vocabulary entries\n", removedWords);
// now prune m-grams against count cut-off
std::vector<int> numMGrams (M+1, 0);
msra::basetypes::fixed_vector<mgram_map::coord> histCoord (M); // index of history mgram
for (int m = 1; m <= M; m++)
{
for (mgram_map::deep_iterator iter (map); iter; ++iter)
{
if (iter.order() != m) continue;
bool prune = counts[iter] < minObs[m-1]; // prune if count below minimum
// prune by vocabulary
const mgram_map::key key = *iter;
for (int k = 0; !prune && k < m; k++)
{
int wid = key[k];
prune |= dropWord[wid];
}
if (prune)
{
counts[iter] = 0; // pruned: this is how we remember it
continue;
}
// for remaining words, check whether the structure is still intact
if (m < M) histCoord[m] = iter;
mgram_map::coord j = histCoord[m-1]; // parent
if (counts[j] == 0)
RuntimeError ("estimate: invalid pruning: a parent m-gram got pruned away");
//throw runtime_error ("estimate: invalid pruning: a parent m-gram got pruned away");
numMGrams[m]++;
}
}
for (int m = 1; m <= M; m++)
{
fprintf (stderr, "estimate: %d-grams after pruning: %d out of %d (%.1f%%)\n", m,
numMGrams[m], counts.size (m),
100.0 * numMGrams[m] / max (counts.size (m), 1));
}
// ensure M reflects the actual order of read data after pruning
while (M > 0 && numMGrams[M] == 0) resize (M-1); // will change M
// === compact memory after pruning
// naw... this is VERY tricky with the mgram_map architecture to keep all data in sync
// So for now we just skip those in all subsequent steps (i.e. we don't save memory)
// === estimate M-gram
fprintf (stderr, "estimate: estimating probabilities...\n");
// dimension the m-gram store
mgram_data<float> P (M); // [M+1][i] probabilities
for (int m = 1; m <= M; m++) P.reserve (m, numMGrams[m]);
// compute discounted probabilities (uninterpolated except, later, for unigram)
// We estimate into a new map so that pruned items get removed.
// For large data sets, where strong pruning is used, there is a net
// memory gain from doing this (we gain if pruning cuts more than half).
mgram_map Pmap (M);
for (int m = 1; m <= M; m++) Pmap.reserve (m, numMGrams[m]);
mgram_map::cache_t PmapCache; // used in map.create()
// m-grams
P.push_back (mgram_map::coord(), 0.0f); // will be updated later
for (int m = 1; m <= M; m++)
{
fprintf (stderr, "estimate: estimating %d %d-gram probabilities...\n", numMGrams[m], m);
// loop over all m-grams of level 'm'
msra::basetypes::fixed_vector<mgram_map::coord> histCoord (m);
for (mgram_map::deep_iterator iter (map, m); iter; ++iter)
{
if (iter.order() != m)
{
// a parent: remember how successors can find their history
// (files are nested like a tree)
histCoord[iter.order()] = iter;
continue;
}
const mgram_map::key key = *iter;
ASSERT (key.order() == iter.order()); // (remove this check once verified)
// get history's count
const mgram_map::coord j = histCoord[m-1]; // index of parent entry
double histCount = counts[j]; // parent count --before pruning
//double histCount = succCount[j]; // parent count --actuals after pruning
// estimate probability for this M-gram
unsigned int count = counts[iter];
// this is 0 for pruned entries
// count = numerator, histCount = denominator
// Kneser-Ney smoothing --replace all but the highest-order
// distribution with that strange Kneser-Ney smoothed distribution.
if (m < M && m < 3 && count > 0) // all non-pruned items except highest order
{ // ^^ seems not to work for 4-gram
// We use a normal distribution if the history is the sentence
// start, as there we fallback without back-off. [Thanks to
// Yining Chen for the tip.]
if (m < 2 || key.pop_w().back() != startId)
{
count = KNCounts[iter]; // (u,v,w) -> count (*,v,w)
if (count == 0) // must exist
RuntimeError ("estimate: malformed data: back-off value not found (numerator)");
const mgram_map::key key_h = key.pop_w();
mgram_map::foundcoord c_h = map[key_h];
if (!c_h.valid_w())
throw runtime_error ("estimate: invalid shortened KN history");
histCount = KNTotalCounts[c_h]; // (u,v,w) -> count (*,v,*)
if (histCount == 0) // must exist
RuntimeError ("estimate: malformed data: back-off value not found (denominator)");
ASSERT (histCount >= count);
}
}
// pruned case
if (count == 0) // this entry was pruned before
goto skippruned;
// <s> does not count as an event, as it is never emitted.
// For now we prune it, but later we put the unigram back with -99.0.
if (key.back() == startId)
{ // (u, v, <s>)
if (m > 1) // do not generate m-grams
goto skippruned;
count = 0; // unigram is kept in structure
}
else if (m == 1)
{ // unigram non-<s> events
histCount--; // do not count <s> in denominator either
// For non-unigrams, we don't need to care because m-gram
// histories of <s> always ends in </s>, and we never ask for such an m-gram
// ... TODO: actually, is subtracting 1 the right thing to do here?
// shouldn't we subtract the unigram count of <s> instead?
}
// Histories with any token before <s> are not valuable, and
// actually need to be removed for consistency with the above
// rule of removing m-grams predicting <s> (if we don't we may
// create orphan m-grams).
for (int k = 1; k < m-1; k++)
{ // ^^ <s> at k=0 and k=m-1 is OK; anywhere else -> useless m-gram
if (key[k] == startId)
goto skippruned;
}
// estimate discounted probability
double dcount = count; // "modified Kneser-Ney" discounting
if (count >= 3) dcount -= d3[m];
else if (count == 2) dcount -= d2[m];
else if (count == 1) dcount -= d1[m];
if (dcount < 0.0) // 0.0 itself is caused by <s>
throw runtime_error ("estimate: negative discounted count value");
if (histCount == 0)
RuntimeError ("estimate: unexpected 0 denominator");
double dP = dcount / histCount;
// and this is the discounted probability value
{
// Actually, 'key' uses a "mapped" word ids, while create()
// expects unmapped ones. However, we have established an
// identical mapping at the start of this function, such that
// we can be sure that key=unmapped key.
mgram_map::coord c = Pmap.create ((mgram_map::unmapped_key) key, PmapCache);
P.push_back (c, (float) dP);
}
skippruned:; // m-gram was pruned
}
}
// the distributions are not normalized --discount mass is missing
fprintf (stderr, "estimate: freeing memory for counts...\n");
KNCounts.clear(); // free some memory
KNTotalCounts.clear();
// the only items used below are P and Pmap.
w2id.resize (Pmap.maxid() +1);
foreach_index (i, w2id) w2id[i] = i;
//std::vector<int> w2id (Pmap.identical_map());
Pmap.created (w2id); // finalize and establish mapping for read access
map.swap (Pmap); // install the new map in our m-gram
Pmap.clear(); // no longer using the old one
counts.clear(); // counts also no longer needed
// zerogram
int vocabSize = 0;
for (mgram_map::iterator iter (map, 1); iter; ++iter)
if (P[iter] > 0.0) // (note: this excludes <s> and all pruned items)
vocabSize++;
P[mgram_map::coord()] = (float) (1.0 / vocabSize); // zerogram probability
// interpolating the unigram with the zerogram
// This is necessary as there is no back-off path from the unigram
// except in the OOV case. I.e. probability mass that was discounted
// from the unigrams is lost. We fix it by using linear interpolation
// instead of strict discounting for the unigram distribution.
double unigramSum = 0.0;
for (mgram_map::iterator iter (map, 1); iter; ++iter)
unigramSum += P[iter];
double missingUnigramMass = 1.0 - unigramSum;
if (missingUnigramMass > 0.0)
{
float missingUnigramProb = (float) (missingUnigramMass * P[mgram_map::coord()]);
fprintf (stderr, "estimate: distributing missing unigram mass of %.2f to %d unigrams\n",
missingUnigramMass, vocabSize);
for (mgram_map::iterator iter (map, 1); iter; ++iter)
{
if (P[iter] == 0.0f) continue; // pruned
P[iter] += missingUnigramProb; // add it in
}
}
// --- M-gram sections --back-off weights
fprintf (stderr, "estimate: determining back-off weights...\n");
computeBackoff (map, M, P, logB, false);
// now the LM is normalized assuming the ARPA back-off computation
// --- take logs and push estimated values into base CMGramLM structure
// take logs in place
for (int m = 0; m <= M; m++)
for (mgram_map::iterator iter (map, m); iter; ++iter)
P[iter] = logclip (P[iter]); // pruned entries go to logzero
P.swap (logP); // swap into base language model
// --- final housekeeping to account for idiosyncrasies of the ARPA format
// resurrect sentence-start symbol with log score -99
const mgram_map::foundcoord cs = map[mgram_map::key (&startId, 1)];
if (cs.valid_w())
logP[cs] = -99.0f * log (10.0f);
// update zerogram prob
// The zerogram will only be used in the OOV case--the non-OOV case has
// been accounted for above by interpolating with the unigram. Thus, we
// replace the zerogram by a value appropriate for an OOV word. We
// choose the minimum unigram prob. This value is not stored in the ARPA
// file, but instead recomputed when loading it. We also reset the
// corresponding back-off weight to 1.0 such that we actually get the
// desired OOV score.
updateOOVScore();
fprintf (stderr, "estimate: done");
for (int m = 1; m <= M; m++) fprintf (stderr, ", %d %d-grams", logP.size (m), m);
fprintf (stderr, "\n");
}
};
// ===========================================================================
// CMGramLMClone -- create CMGramLM from sub-LMs through ILM and ILM::IIter
// - create in memory into a CMGramLM
// - write to ARPA file (static function)
// ===========================================================================
class CMGramLMClone : public CMGramLM
{
public:
// create an LM in memory iterating through an underlying model
// This uses IILM::IIter and the score() function, i.e. it works for all
// derivative LM types such as linear interpolation.
// Back-off weights are recomputed in this function. I.e. even if applied
// to a plain m-gram, results may be different if tricks were played with
// the back-off weights in the original model.
// The dropWord[] vector, if not empty, specifies a set of words that
// should be dropped (m-grams that contain such a word are skipped).
// Underlying models are assumed to have m-gram property, otherwise the
// resulting LM will explode.
void clone (const ILM & lm, int p_M = INT_MAX, const vector<bool> & dropWord = vector<bool>())
{
if (p_M > lm.order())
p_M = lm.order();
M = p_M;
map.init (M);
logP.init (M);
logB.init (M-1);
// allocate the memory
for (int m = 0; m <= M; m++)
{
size_t size_m = lm.size (m);
map.reserve (m, size_m);
logP.reserve (m, size_m);
if (m < M)
logB.reserve (m, size_m);
}
// compute the scores
// Iterator will iterate in increasing order of word ids as returned
// by *iter.
bool filterWords = !dropWord.empty();
mgram_map::cache_t mapCache;
auto_ptr<IIter> piter (lm.iter (0, M));
for (IIter & iter = *piter; iter; ++iter)
{
// get key (mgram[], m) for current iter position
std::pair<const int*,int> keyp = *iter;
const int * mgram = keyp.first;
int m = keyp.second;
mgram_map::unmapped_key key (mgram, m);
// skip if we filter against a dropWord[] list
if (filterWords)
{
// if any of the dropWord[mgram[]] is set then skip
for (int i = 0; i < key.order(); i++)
{
int w = key[i];
if (dropWord[w])
goto skipMGram;//skipMGram
}
}
// local block for get rid of: warning C4533: initialization of 'c' is skipped by 'goto skipMGram'
// (local block because we 'goto' over this)
{
// create map entry
mgram_map::coord c = map.create (key, mapCache);
// create probability entry
double thisLogP = lm.score (mgram, m);
logP.push_back (c, (float) thisLogP);
}
skipMGram:
filterWords = filterWords;
}
// finalize map and establish w->id mapping (identical)
std::vector<int> w2id (map.identical_map());
map.created (w2id);
// create back-off data
computeBackoff (map, M, logP, logB, true);
// and replace zerogram score by the OOV score
updateOOVScore();
}
// static function to clone a model and write it out as an ARPA (text) file.
// Symbols can be anything that has symbols[w] -> std::string& .
// A future version may do this more efficiently.
template<class SYMMAP>
static void write (const ILM & lm, int M, FILE * outf, const SYMMAP & symbols)
{
fprintf (stderr, "write: cloning...\n");
CMGramLMClone outlm;
outlm.clone (lm, M);
fprintf (stderr, "write: saving...\n");
((const CMGramLM&) outlm).write (outf, symbols);
}
// read and parse a #clone file
static void read (const wstring & clonepath, wstring & lmpath)
{
wstring dir, file;
splitpath (clonepath, dir, file); // we allow relative paths in the file
msra::basetypes::auto_file_ptr f = fopenOrDie (clonepath, L"rbS");
std::string line = fgetline (f);
if (line != "#clone")
throw runtime_error ("read: invalid header line " + line);
std::string lmpath8 = fgetline (f); // only one item: the pathname
if (lmpath8.empty())
throw runtime_error ("read: pathname missing");
lmpath = msra::strfun::utf16 (lmpath8);
}
};
#if 0 // old version --remove once we are fully tested and comfortable
class OldCMGramLM : public ILM
{
protected:
// representation of LM in memory
// For each order, there is a flattened array of LMSCORE tokens.
// For each history order, there is a flattened array of LMHISTs.
// E.g. a trigram's history's LMHIST entry (somewhere in refs[2]) denotes
// the start index of the first LMSCORE entry (in entries[3]). The end
// index is denoted by the start index of the next LMHIST entry (for this
// purpose, the LMHIST arrays have one extra entry at the end).
struct LMSCORE // an LM score, plus its word id for sparse storage
{
int id; // token id (in LM space)
float logP; // and its score
LMSCORE (int p_id, double p_logP) : id (p_id), logP ((float) p_logP) { }
};
struct LMHIST // an LM history -- index corresponds to LMSCORE index
{
int firstEntry; // index of first entry (end entry known from next LMHIST)
float logB; // back-off weight
LMHIST (int p_firstEntry, double p_logB) : firstEntry (p_firstEntry), logB ((float) p_logB) { }
};
int M;
std::vector<std::vector<LMHIST>> refs; // [M] e.g. [2] for trigram history
std::vector<std::vector<LMSCORE>> entries; // [M+1] e.g. [3] for trigrams. [0]=dummy
// mapping of numeric word ids from external (user-defined) space to the internal LM's
std::vector<int> userToLMSymMap; // map to ids used in LM
// map user id to LM id, return -1 for anything unknown
inline int mapId (int userId) const
{
if (userId < 0 || userId >= (int) userToLMSymMap.size()) return -1;
else return userToLMSymMap[userId];
}
bool entries1Unmapped; // if true then findEntry(id) == i for entries[1]
// search in an LMSCORE array
// This is a relatively generic binary search.
inline int findEntry (const std::vector<LMSCORE> & entries, int beg, int end, int id) const
{
while (beg < end)
{
int i = (beg + end) / 2;
int v = entries[i].id;
if (id == v) return i; // found it
else if (id < v) end = i; // id is left of i
else beg = i + 1; // id is right of i
}
return -1; // not found
}
// diagnostics of previous score() call
mutable int longestMGramFound; // longest m-gram (incl. predicted token) found
mutable int longestHistoryFound; // longest history (excl. predicted token) found
public:
virtual int getLastLongestHistoryFound() const { return longestHistoryFound; }
virtual int getLastLongestMGramFound() const { return longestMGramFound; }
virtual int order() const { return M; }
// mgram[m-1] = word to predict, tokens before that are history
// m=3 means trigram
virtual double score (const int * mgram, int m) const
{
longestHistoryFound = 0; // (diagnostics)
if (m > M) // too long a history for this model
{
mgram += (m - M);
m = M;
}
double totalLogB = 0.0; // accumulated back-off
for (;;)
{
longestMGramFound = m; // (diagnostics)
if (m == 0) // not really defined in ARPA format
return totalLogB + entries[0][0].logP;
if (m == 1)
{
// find the actual score
// [beg, end) is the sub-range in entries array.
int id = mapId (mgram[0]);
const char * sym = idToSymbol (id); sym;// (debugging)
const std::vector<LMSCORE> & entries_1 = entries[1];
int i = entries1Unmapped ? id : findEntry (entries_1, refs[0][0].firstEntry, refs[0][1].firstEntry, id);
if (i == -1)
goto backoff0;
ASSERT (entries_1[i].id == id); // verify unmapped unigram case
double logP = entries_1[i].logP;
return totalLogB + logP;
}
// locate LMHIST and LMSCORE
// We traverse history one by one.
int id = mapId (mgram[0]); // start with unigram history
const char * sym = idToSymbol (id); // (debugging)
int i = (entries1Unmapped) ? id : findEntry (entries[1], refs[0][0].firstEntry, refs[0][1].firstEntry, id);
if (i == -1) // unknown history: fall back
goto fallback;
ASSERT (entries[1][i].id == id); // verify unmapped unigram case
// found it: advance search by one history token
const std::vector<LMHIST> & refs_1 = refs[1];
float logB = refs_1[i].logB;
int beg = refs_1[i].firstEntry; // sub-array range for next level
int end = refs_1[i+1].firstEntry;
for (int n = 2; n < m; n++)
{
if (beg == end)
goto fallback; // unseen history: fall back
int id = mapId (mgram[n -1]);
const char * sym = idToSymbol (id); sym; // (debugging)
int i = findEntry (entries[n], beg, end, id);
if (i == -1) // unseen history: fall back
goto fallback;
ASSERT (entries[n][i].id == id); // verify unmapped unigram case
// found it: advance search by one history token
const std::vector<LMHIST> & refs_n = refs[n];
logB = refs_n[i].logB;
beg = refs_n[i].firstEntry; // sub-array range for next level
end = refs_n[i+1].firstEntry;
}
// we found the entire history: now find the actual score
// [beg, end) is the sub-range in entries array.
if (m -1 > longestHistoryFound)
longestHistoryFound = m -1;
if (beg == end) // history has no successors (but a back-off weight)
goto backoff;
id = mapId (mgram[m -1]);
sym = idToSymbol (id); // (debugging)
const std::vector<LMSCORE> & entries_m = entries[m];
i = findEntry (entries_m, beg, end, id);
if (i == -1)
goto backoff;
ASSERT (entries_m[i].id == id); // verify unmapped unigram case
longestMGramFound = m;
double logP = entries_m[i].logP;
return totalLogB + logP;
backoff: // found history but not predicted token: back-off
totalLogB += logB;
backoff0: // back-off knowing that logB == 0
fallback: // we get here in case of fallback (no back-off weight) or back-off
mgram++;
m--;
} // and go again with the shortened history
}
// same as score() but without optimizations (for reference)
double score_unoptimized (const int * mgram, int m) const
{
if (m == 0) // not really defined in ARPA format
return entries[0][0].logP;
else if (m > M) // too long a history for this model
{
mgram += (m - M);
m = M;
}
// locate LMHIST and LMSCORE
// We traverse history one by one.
int beg = refs[0][0].firstEntry; // start with the unigram array
int end = refs[0][1].firstEntry;
float logB = 0.0f; // remember in the loop in case we need it
for (int n = 1; n < m; n++)
{
int userId = mgram[n -1]; // may be -1 for unknown word
int id = mapId (userId);
const char * sym = idToSymbol (id); sym; // (debugging)
const std::vector<LMSCORE> & entries_n = entries[n];
int i = findEntry (entries_n, beg, end, id);
if (i == -1) // unknown history: fall back
return score_unoptimized (mgram +1, m -1); // tail recursion
ASSERT (entries_n[i].id == id); // verify unmapped unigram case
// found it: advance search by one history token
const std::vector<LMHIST> & refs_n = refs[n];
logB = refs_n[i].logB;
beg = refs_n[i].firstEntry; // sub-array range for next level
end = refs_n[i+1].firstEntry;
}
// we found the entire history: now find the actual score
// [beg, end) is the sub-range in entries array.
int userId = mgram[m -1]; // word to predict
int id = mapId (userId);
const char * sym = idToSymbol (id); sym; // (debugging)
const std::vector<LMSCORE> & entries_m1 = entries[m];
int i = findEntry (entries_m1, beg, end, id);
if (i != -1)
{
ASSERT (entries_m1[i].id == id); // verify unmapped unigram case
double logP = entries_m1[i].logP;
return logP;
}
// found history but not predicted token: back-off
return logB + score_unoptimized (mgram + 1, m -1);
}
// test for OOV word (OOV w.r.t. LM)
virtual bool oov (int id) const { return mapId (id) < 0; }
virtual void adapt (const int *, size_t) { } // this LM does not adapt
private:
// keep this for debugging
std::wstring filename; // input filename
struct SYMBOL
{
string symbol; // token
int id; // numeric id in LM space (index of word read)
bool operator< (const SYMBOL & other) const { return symbol < other.symbol; }
SYMBOL (int p_id, const char * p_symbol) : id (p_id), symbol (p_symbol) { }
};
std::vector<SYMBOL> lmSymbols; // (id, word) symbols used in LM
std::vector<int> idToSymIndex; // map LM id to index in lmSymbols[] array
// search for a word in the sorted word array.
// Only use this after sorting, i.e. after full 1-gram section has been read.
// Only really used in read().
inline int symbolToId (const char * word) const
{
int beg = 0;
int end = (int) lmSymbols.size();
while (beg < end)
{
int i = (beg + end) / 2;
const char * v = lmSymbols[i].symbol.c_str();
int cmp = strcmp (word, v);
if (cmp == 0) return lmSymbols[i].id; // found it
else if (cmp < 0) end = i; // id is left of i
else beg = i + 1; // id is right of i
}
return -1; // not found
}
inline const char * idToSymbol (int id) const
{
if (id < 0) return NULL; // empty string for unknown ids
int i = idToSymIndex[id];
return lmSymbols[i].symbol.c_str();
}
public:
// read an ARPA (text) file.
// Words do not need to be sorted in the unigram section, but the m-gram
// sections have to be in the same order as the unigrams.
// The 'userSymMap' defines the vocabulary space used in score().
// If 'filterVocabulary' then LM entries for words not in userSymMap are skipped.
// Otherwise the userSymMap is updated with the words from the LM.
// 'maxM' allows to restrict the loading to a smaller LM order.
// SYMMAP can be e.g. CSymMap or CSymbolSet.
template<class SYMMAP>
void read (const std::wstring & pathname, SYMMAP & userSymMap, bool filterVocabulary, int maxM)
{
int lineNo = 0;
msra::basetypes::auto_file_ptr f = fopenOrDie (pathname, L"rbS");
fprintf (stderr, "read: reading %S", pathname.c_str());
filename = pathname; // (keep this info for debugging)
// --- read header information
// search for header line
char buf[1024];
lineNo++, fgetline (f, buf);
while (strcmp (buf, "\\data\\") != 0 && !feof (f))
lineNo++, fgetline (f, buf);
lineNo++, fgetline (f, buf);
// get the dimensions
std::vector<int> dims; dims.reserve (4);
while (buf[0] == 0 && !feof (f))
lineNo++, fgetline (f, buf);
int n, dim;
dims.push_back (1); // dummy zerogram entry
while (sscanf (buf, "ngram %d=%d", &n, &dim) == 2 && n == (int) dims.size())
{
dims.push_back (dim);
lineNo++, fgetline (f, buf);
}
M = (int) dims.size() -1;
if (M == 0)
RuntimeError ("read: mal-formed LM file, no dimension information (%d): %S", lineNo, pathname.c_str());
int fileM = M;
if (M > maxM)
M = maxM;
// allocate main storage
refs.resize (M);
for (int m = 0; m < M; m++)
refs[m].reserve (dims[m] +1);
entries.resize (M +1);
for (int m = 0; m <= M; m++)
entries[m].reserve (dims[m]);
lmSymbols.reserve (dims[0]);
refs[0].push_back (LMHIST (0, 0.0));
refs[0].push_back (LMHIST (0, -99.0)); // this one gets updated
entries[0].push_back (LMSCORE (-1, -99.0)); // zerogram score -- gets updated later
std::vector<bool> skipWord; // true: skip entry containing this word
skipWord.reserve (lmSymbols.capacity());
// --- read main sections
const double ln10xLMF = log (10.0); // ARPA scores are strangely scaled
for (int m = 1; m <= M; m++)
{
while (buf[0] == 0 && !feof (f))
lineNo++, fgetline (f, buf);
if (sscanf (buf, "\\%d-grams:", &n) != 1 || n != m)
RuntimeError ("read: mal-formed LM file, bad section header (%d): %S", lineNo, pathname.c_str());
lineNo++, fgetline (f, buf);
std::vector<int> mgram (m +1); // current mgram being read
std::vector<int> prevmgram (m +1, -1); // previous mgram read
std::vector<int> histEntry (m); // sub-array ranges
histEntry[0] = 0;
// read all the m-grams
while (buf[0] != '\\')
{
if (buf[0] == 0)
{
lineNo++, fgetline (f, buf);
continue;
}
// -- parse the line
const char * delim = " \t\n\r";
const char * score = strtok (&buf[0], delim);
if (score == NULL || score[0] == 0) // not checking whether it is numeric
RuntimeError ("read: mal-formed LM file, no score (%d): %S", lineNo, pathname.c_str());
double scoreVal = atof (score);
double logP = scoreVal * ln10xLMF; // convert to natural log
bool skipEntry = false;
for (int n = 1; n <= m; n++)
{
/*const*/ char * tok = strtok (NULL, delim);
if (tok == NULL)
RuntimeError ("read: mal-formed LM file, not enough words in mgram (%d): %S", lineNo, pathname.c_str());
// map to id
int id;
if (m == 1) // unigram: build vocab table
{
id = (int) lmSymbols.size(); // unique id for this symbol
lmSymbols.push_back (SYMBOL (id, tok));
bool toSkip = false;
if (userSymMap.sym2existingId (lmSymbols.back().symbol) == -1)
{
if (filterVocabulary)
toSkip = true; // unknown word
else
userSymMap.sym2id (lmSymbols.back().symbol); // create it in user's space
}
skipWord.push_back (toSkip);
}
else // mgram: look up word in vocabulary
{
if (prevmgram[n] >= 0 && strcmp (idToSymbol (prevmgram[n]), tok) == 0)
id = prevmgram[n];
else
{
id = symbolToId (tok);
if (id == -1)
RuntimeError ("read: mal-formed LM file, m-gram contains unknown word (%d): %S", lineNo, pathname.c_str());
}
}
mgram[n] = id; // that's our id
skipEntry |= skipWord[id]; // skip entry if any token is unknown
}
double logB = 0.0;
if (m < M)
{
const char * bo = strtok (NULL, delim);
if (score == NULL || score[0] == 0) // not checking whether it is numeric
RuntimeError ("read: mal-formed LM file, no score (%d): %S", lineNo, pathname.c_str());
double boVal = atof (bo);
logB = boVal * ln10xLMF; // convert to natural log
}
lineNo++, fgetline (f, buf);
if (skipEntry) // word contained unknown vocabulary: skip entire entry
goto skipMGram;
// -- enter the information into our data structure
// locate the corresponding entries
// histEntry[n] are valid iff mgram[n'] == prevmgram[n'] for all n' <= '
bool prevValid = true;
for (int n = 1; n < m; n++)
{
if (prevValid && mgram[n] == prevmgram[n])
continue;
if (prevValid && mgram[n] < prevmgram[n])
RuntimeError ("read: mal-formed LM file, m-gram out of order (%d): %S", lineNo, pathname.c_str());
// a history token differs from previous mgram. That history must exist.
const std::vector<LMSCORE> & entries_n = entries[n];
const std::vector<LMHIST> & refs_h = refs[n -1]; // history
int beg = refs_h[histEntry[n -1]].firstEntry; // sub-array range for next level
int end = refs_h[histEntry[n -1] +1].firstEntry;
int i = findEntry (entries_n, beg, end, mgram[n]);
if (i == -1) // unknown history: fall back
RuntimeError ("read: mal-formed LM file, m-gram history not defined (%d): %S", lineNo, pathname.c_str());
// found it: narrow down search range
histEntry[n] = i;
prevValid = false;
}
if (prevValid && mgram[m] <= prevmgram[m])
RuntimeError ("read: mal-formed LM file, m-gram out of order (%d): %S", lineNo, pathname.c_str());
if (m < M) // create history entry
refs[m].push_back (LMHIST (0, logB));
entries[m].push_back (LMSCORE (mgram[m], logP)); // score entry
refs[m-1][histEntry[m-1]].firstEntry++; // for now count how many histories we got
skipMGram:
// remember current mgram for next iteration
::swap (mgram, prevmgram);
}
// Update previous level history from #entries to firstEntry.
// We do this afterwards because some histories may not be used and
// therefore not occur in higher-order m-grams, such that we cannot
// rely on touching them in the loop above. Counting entries instead
// leaves those at 0, which is correct.
std::vector<LMHIST> & refs_h = refs[m -1]; // history
int n0 = 0;
for (int i = 0; i < (int) refs_h.size(); i++)
{
int num = refs_h[i].firstEntry;
refs_h[i].firstEntry = n0;
n0 += num;
}
ASSERT (refs_h.back().firstEntry == (int) entries[m].size());
// create closing history entry
if (m < M)
refs[m].push_back (LMHIST (0, -99.0));
// fix the symbol set -- now we can binary-search in them with symbolToId()
if (m == 1)
{
std::sort (lmSymbols.begin(), lmSymbols.end());
idToSymIndex.resize (lmSymbols.size(), -1);
for (int i = 0; i < (int) lmSymbols.size(); i++)
{
idToSymIndex[lmSymbols[i].id] = i;
}
}
fprintf (stderr, ", %d %d-grams", entries[m].size(), m);
}
fprintf (stderr, "\n");
// check end tag
if (M == fileM)
{ // only if caller did not restrict us to a lower order
while (buf[0] == 0 && !feof (f))
lineNo++, fgetline (f, buf);
if (strcmp (buf, "\\end\\") != 0)
RuntimeError ("read: mal-formed LM file, no \\end\\ tag (%d): %S", lineNo, pathname.c_str());
}
// update zerogram score
// We use the minimum of all unigram scores.
const std::vector<LMSCORE> & entries_1 = entries[1];
float unknownLogP = 0.0f;
for (int i = 0; i < (int) entries_1.size(); i++)
{
if (entries_1[i].logP < -98.9f) continue; // disabled token does not count
if (entries_1[i].logP < unknownLogP)
unknownLogP = entries_1[i].logP;
}
entries[0][0].logP = unknownLogP;;
//= (float) -log ((double) lmSymbols.size()); // zerogram score
// establish mapping of word ids from user to LM space
userToLMSymMap.resize (userSymMap.size());
for (int i = 0; i < userSymMap.size(); i++)
{
const char * sym = userSymMap.id2sym (i);
int id = symbolToId (sym); // may be -1 if not found
userToLMSymMap[i] = id;
}
// check whether first-level unigrams need mapping
// We don't unless user provided a dictionary to filter.
entries1Unmapped = true; // assume findEntry (id) == id
for (int i = 0; i < (int) entries_1.size(); i++)
{
if (entries_1[i].id != i)
{
entries1Unmapped = false;
break;
}
}
}
};
#endif
// ===========================================================================
// CPerplexity -- helper to measure perplexity
// ===========================================================================
class CPerplexity
{
double logPAcc; // accumulated logP
int numTokensAcc; // tokens accumulated
int numOOVTokens; // OOVs have been skipped
int numUtterances;
const ILM & lm;
int startId, endId;
std::vector<int> buf; // temp buffer to insert <s> and </s>
CPerplexity & operator= (const CPerplexity &); // inaccessible
public:
CPerplexity (const ILM & p_lm, int p_startId, int p_endId) : lm (p_lm), startId (p_startId), endId (p_endId)
{ buf.reserve (1000); reset(); }
// reset perplexity accumulation (clear all that's been passed)
void reset() { logPAcc = 0.0; numTokensAcc = numOOVTokens = numUtterances = 0; }
// Add perplexity for an utterance. Ids are in the same numeric id
// space as was used to read() the language model. Only the actual words
// should be included in ids[], do not include sentence start/end markers.
// These are implied by this function.
template<class SYMMAP>
void addUtterance (const std::vector<int> & ids, const SYMMAP & symMap)
{
buf.assign (1, startId);
buf.insert (buf.end(), ids.begin(), ids.end());
buf.push_back (endId);
for (int i = 1; i < (int) buf.size(); i++)
{
if (lm.oov (buf[i])) // silently skip words unknown to the LM
{
numOOVTokens++;
continue;
}
double logP = lm.score (&buf[0], i +1); // use full history
if (logP <= -1e20)
{
#if 0 // should really not happen
fprintf (stderr, "skipping poor-scoring %s (%.2f)\n", symMap[buf[i]], logP);
#endif
numOOVTokens++;
continue;
}
#if 0 // analysis of back-off etc.
// dump some interesting information
int mseenhist = lm.getLastLongestHistoryFound();
int mseen = lm.getLastLongestMGramFound();
int order = lm.order();
if (order > i+1) // limit to what we've requested
order = i+1;
char pbuf[20];
sprintf (pbuf, "%7.5f", exp (logP));
for (int k = 2; pbuf[k]; k++) if (pbuf[k] == '0') pbuf[k] = '.'; else break;
char smseenhist[20]; // fallback=length of actual history
smseenhist[order-1] = 0;
for (int k = 0; k < order -1; k++) smseenhist[k] = (k >= order-1 - mseenhist) ? '.' : 'X';
char smseen[20];
smseen[order] = 0;
for (int k = 0; k < order; k++) smseen[k] = (k >= order - mseen) ? '.' : 'X';
char seq[100] = { 0 };
for (int i1 = i - (order-1); i1 <= i; i1++)
{
strcat (seq, "_");
strcat (seq, symMap[buf[i1]]);
}
fprintf (stderr, "=%-22s\t%6.2f\t%s\t%s %s\n", seq+1, logP, pbuf +1, smseenhist, smseen);
#else
symMap;
#endif
#if 0 // testing of optimization
double logP1 = lm.score_unoptimized (&buf[0], i +1); // use full history
if (fabs (logP - logP1) > 1e-3)
RuntimeError ("bug in optimized score()");
#endif
logPAcc += logP;
numTokensAcc++;
}
numUtterances++;
}
// return perplexity of words accumulated so far
double getPerplexity() const
{
double avLogP = logPAcc / max (numTokensAcc, 1);
double PPL = exp (-avLogP);
return PPL;
}
// return number of words passed in, including OOV tokens (but not implied sentence ends)
int getNumWords() const { return numTokensAcc + numOOVTokens - numUtterances; }
// return number of OOV tokens
int getNumOOV() const { return numOOVTokens; }
// return number of utterances
int getNumUtterances() const { return numUtterances; }
};
};}; // namespace