git-svn-id: https://svn.code.sf.net/p/kaldi/code/trunk@716 5e6a8d80-dfce-4ca6-a32a-6e07a63d50c8
This commit is contained in:
Dan Povey 2012-02-07 15:06:21 +00:00
Родитель 702ab77b30
Коммит 08274c12a2
1 изменённых файлов: 225 добавлений и 0 удалений

225
tools/openfst.patch Normal file
Просмотреть файл

@ -0,0 +1,225 @@
*** partition.h
***************
*** 41,48 ****
friend class PartitionIterator<T>;
struct Element {
! Element() : value(0), next(0), prev(0) {}
! Element(T v) : value(v), next(0), prev(0) {}
T value;
Element* next;
--- 59,66 ----
friend class PartitionIterator<T>;
struct Element {
! Element() : value(0), next(0), prev(0) {}
! Element(T v) : value(v), next(0), prev(0) {}
T value;
Element* next;
***************
*** 50,58 ****
};
public:
! Partition() {}
! Partition(T num_states) {
Initialize(num_states);
}
--- 68,78 ----
};
public:
! Partition(bool allow_repeated_split):
! allow_repeated_split_(allow_repeated_split) {}
! Partition(bool allow_repeated_split, T num_states):
! allow_repeated_split_(allow_repeated_split) {
Initialize(num_states);
}
***************
*** 135,150 ****
if (class_size_[class_id] == 1) return;
// first time class is split
! if (split_size_[class_id] == 0)
visited_classes_.push_back(class_id);
!
// increment size of split (set of element at head of chain)
split_size_[class_id]++;
!
// update split point
! if (class_split_[class_id] == 0)
! class_split_[class_id] = classes_[class_id];
! if (class_split_[class_id] == elements_[element_id])
class_split_[class_id] = elements_[element_id]->next;
// move to head of chain in same class
--- 155,170 ----
if (class_size_[class_id] == 1) return;
// first time class is split
! if (split_size_[class_id] == 0) {
visited_classes_.push_back(class_id);
! class_split_[class_id] = classes_[class_id];
! }
// increment size of split (set of element at head of chain)
split_size_[class_id]++;
!
// update split point
! if (class_split_[class_id] != 0
! && class_split_[class_id] == elements_[element_id])
class_split_[class_id] = elements_[element_id]->next;
// move to head of chain in same class
***************
*** 155,178 ****
// class indices of the newly created class. Returns the new_class id
// or -1 if no new class was created.
T SplitRefine(T class_id) {
// only split if necessary
! if (class_size_[class_id] == split_size_[class_id]) {
! class_split_[class_id] = 0;
split_size_[class_id] = 0;
return -1;
} else {
-
T new_class = AddClass();
size_t remainder = class_size_[class_id] - split_size_[class_id];
if (remainder < split_size_[class_id]) { // add smaller
- Element* split_el = class_split_[class_id];
classes_[new_class] = split_el;
- class_size_[class_id] = split_size_[class_id];
- class_size_[new_class] = remainder;
split_el->prev->next = 0;
split_el->prev = 0;
} else {
- Element* split_el = class_split_[class_id];
classes_[new_class] = classes_[class_id];
class_size_[class_id] = remainder;
class_size_[new_class] = split_size_[class_id];
--- 175,205 ----
// class indices of the newly created class. Returns the new_class id
// or -1 if no new class was created.
T SplitRefine(T class_id) {
+
+ Element* split_el = class_split_[class_id];
// only split if necessary
! //if (class_size_[class_id] == split_size_[class_id]) {
! if(split_el == NULL) { // we split on everything...
split_size_[class_id] = 0;
return -1;
} else {
T new_class = AddClass();
+
+ if(allow_repeated_split_) { // split_size_ is possibly
+ // inaccurate, so work it out exactly.
+ size_t split_count; Element *e;
+ for(split_count=0,e=classes_[class_id];
+ e != split_el; split_count++, e=e->next);
+ split_size_[class_id] = split_count;
+ }
size_t remainder = class_size_[class_id] - split_size_[class_id];
if (remainder < split_size_[class_id]) { // add smaller
classes_[new_class] = split_el;
split_el->prev->next = 0;
split_el->prev = 0;
+ class_size_[class_id] = split_size_[class_id];
+ class_size_[new_class] = remainder;
} else {
classes_[new_class] = classes_[class_id];
class_size_[class_id] = remainder;
class_size_[new_class] = split_size_[class_id];
***************
*** 243,252 ****
--- 270,285 ----
vector<T> class_size_;
// size of split for each class
+ // in the nondeterministic case, split_size_ is actually an upper
+ // bound on the size of split for each class.
vector<T> split_size_;
// set of visited classes to be used in split refine
vector<T> visited_classes_;
+
+ // true if input fst was deterministic: we can make
+ // certain assumptions in this case that speed up the algorithm.
+ bool allow_repeated_split_;
};
*** minimize.h
***************
*** 133,139 ****
typedef typename A::Weight Weight;
typedef ReverseArc<A> RevA;
! CyclicMinimizer(const ExpandedFst<A>& fst) {
Initialize(fst);
Compute(fst);
}
--- 150,163 ----
typedef typename A::Weight Weight;
typedef ReverseArc<A> RevA;
! CyclicMinimizer(const ExpandedFst<A>& fst):
! // tell the Partition data-member to expect multiple repeated
! // calls to SplitOn with the same element if we are non-deterministic.
! P_(fst.Properties(kIDeterministic, true) == 0) {
! if(fst.Properties(kIDeterministic, true) == 0)
! CHECK(Weight::Properties() & kIdempotent); // this minimization
! // algorithm for non-deterministic FSTs can only work with idempotent
! // semirings.
Initialize(fst);
Compute(fst);
}
***************
*** 314,320 ****
typedef typename A::StateId ClassId;
typedef typename A::Weight Weight;
! AcyclicMinimizer(const ExpandedFst<A>& fst) {
Initialize(fst);
Refine(fst);
}
--- 338,350 ----
typedef typename A::StateId ClassId;
typedef typename A::Weight Weight;
! AcyclicMinimizer(const ExpandedFst<A>& fst):
! // tell the Partition data-member to expect multiple repeated
! // calls to SplitOn with the same element if we are non-deterministic.
! partition_(fst.Properties(kIDeterministic, true) == 0) {
! if(fst.Properties(kIDeterministic, true) == 0)
! CHECK(Weight::Properties() & kIdempotent); // minimization for
! // non-deterministic FSTs can only work with idempotent semirings.
Initialize(fst);
Refine(fst);
}
***************
*** 528,538 ****
void Minimize(MutableFst<A>* fst,
MutableFst<A>* sfst = 0,
float delta = kDelta) {
! uint64 props = fst->Properties(kAcceptor | kIDeterministic|
! kWeighted | kUnweighted, true);
! if (!(props & kIDeterministic))
! LOG(FATAL) << "Input Fst is not deterministic";
!
if (!(props & kAcceptor)) { // weighted transducer
VectorFst< GallicArc<A, STRING_LEFT> > gfst;
Map(*fst, &gfst, ToGallicMapper<A, STRING_LEFT>());
--- 558,565 ----
void Minimize(MutableFst<A>* fst,
MutableFst<A>* sfst = 0,
float delta = kDelta) {
! uint64 props = fst->Properties(kAcceptor | kWeighted | kUnweighted, true);
!
if (!(props & kAcceptor)) { // weighted transducer
VectorFst< GallicArc<A, STRING_LEFT> > gfst;
Map(*fst, &gfst, ToGallicMapper<A, STRING_LEFT>());