提交 ad8ec177 编写于 作者: Y Yang Zhou

add tlg decoder

上级 b5315657
...@@ -3,3 +3,7 @@ cmake_minimum_required(VERSION 3.14 FATAL_ERROR) ...@@ -3,3 +3,7 @@ cmake_minimum_required(VERSION 3.14 FATAL_ERROR)
add_executable(offline_decoder_main ${CMAKE_CURRENT_SOURCE_DIR}/offline_decoder_main.cc) add_executable(offline_decoder_main ${CMAKE_CURRENT_SOURCE_DIR}/offline_decoder_main.cc)
target_include_directories(offline_decoder_main PRIVATE ${SPEECHX_ROOT} ${SPEECHX_ROOT}/kaldi) target_include_directories(offline_decoder_main PRIVATE ${SPEECHX_ROOT} ${SPEECHX_ROOT}/kaldi)
target_link_libraries(offline_decoder_main PUBLIC nnet decoder fst utils gflags glog kaldi-base kaldi-matrix kaldi-util ${DEPS}) target_link_libraries(offline_decoder_main PUBLIC nnet decoder fst utils gflags glog kaldi-base kaldi-matrix kaldi-util ${DEPS})
add_executable(offline_wfst_decoder_main ${CMAKE_CURRENT_SOURCE_DIR}/offline_wfst_decoder_main.cc)
target_include_directories(offline_wfst_decoder_main PRIVATE ${SPEECHX_ROOT} ${SPEECHX_ROOT}/kaldi)
target_link_libraries(offline_wfst_decoder_main PUBLIC nnet decoder fst utils gflags glog kaldi-base kaldi-matrix kaldi-util kaldi-decoder ${DEPS})
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
// todo refactor, repalce with gtest
#include "base/flags.h"
#include "base/log.h"
#include "decoder/ctc_tlg_decoder.h"
#include "frontend/raw_audio.h"
#include "kaldi/util/table-types.h"
#include "nnet/decodable.h"
#include "nnet/paddle_nnet.h"
DEFINE_string(feature_respecifier, "", "test feature rspecifier");
DEFINE_string(model_path, "avg_1.jit.pdmodel", "paddle nnet model");
DEFINE_string(param_path, "avg_1.jit.pdiparams", "paddle nnet model param");
DEFINE_string(word_symbol_table, "vocab.txt", "word symbol table");
DEFINE_string(graph_path, "TLG", "decoder graph");
using kaldi::BaseFloat;
using kaldi::Matrix;
using std::vector;
int main(int argc, char* argv[]) {
gflags::ParseCommandLineFlags(&argc, &argv, false);
google::InitGoogleLogging(argv[0]);
kaldi::SequentialBaseFloatMatrixReader feature_reader(
FLAGS_feature_respecifier);
std::string model_graph = FLAGS_model_path;
std::string model_params = FLAGS_param_path;
std::string word_symbol_table = FLAGS_word_symbol_table;
std::string graph_path = FLAGS_graph_path;
int32 num_done = 0, num_err = 0;
ppspeech::TLGDecoderOptions opts;
opts.word_symbol_table = word_symbol_table;
opts.fst_path = graph_path;
ppspeech::TLGDecoder decoder(opts);
ppspeech::ModelOptions model_opts;
model_opts.model_path = model_graph;
model_opts.params_path = model_params;
std::shared_ptr<ppspeech::PaddleNnet> nnet(
new ppspeech::PaddleNnet(model_opts));
std::shared_ptr<ppspeech::RawDataCache> raw_data(
new ppspeech::RawDataCache());
std::shared_ptr<ppspeech::Decodable> decodable(
new ppspeech::Decodable(nnet, raw_data));
int32 chunk_size = 35;
decoder.InitDecoder();
for (; !feature_reader.Done(); feature_reader.Next()) {
string utt = feature_reader.Key();
const kaldi::Matrix<BaseFloat> feature = feature_reader.Value();
raw_data->SetDim(feature.NumCols());
int32 row_idx = 0;
int32 num_chunks = feature.NumRows() / chunk_size;
for (int chunk_idx = 0; chunk_idx < num_chunks; ++chunk_idx) {
kaldi::Vector<kaldi::BaseFloat> feature_chunk(chunk_size *
feature.NumCols());
for (int row_id = 0; row_id < chunk_size; ++row_id) {
kaldi::SubVector<kaldi::BaseFloat> tmp(feature, row_idx);
kaldi::SubVector<kaldi::BaseFloat> f_chunk_tmp(
feature_chunk.Data() + row_id * feature.NumCols(),
feature.NumCols());
f_chunk_tmp.CopyFromVec(tmp);
row_idx++;
}
raw_data->Accept(feature_chunk);
if (chunk_idx == num_chunks - 1) {
raw_data->SetFinished();
}
decoder.AdvanceDecode(decodable);
}
std::string result;
result = decoder.GetFinalBestPath();
KALDI_LOG << " the result of " << utt << " is " << result;
decodable->Reset();
decoder.Reset();
++num_done;
}
KALDI_LOG << "Done " << num_done << " utterances, " << num_err
<< " with errors.";
return (num_done != 0 ? 0 : 1);
}
...@@ -6,5 +6,6 @@ add_library(decoder STATIC ...@@ -6,5 +6,6 @@ add_library(decoder STATIC
ctc_decoders/decoder_utils.cpp ctc_decoders/decoder_utils.cpp
ctc_decoders/path_trie.cpp ctc_decoders/path_trie.cpp
ctc_decoders/scorer.cpp ctc_decoders/scorer.cpp
ctc_tlg_decoder.cc
) )
target_link_libraries(decoder PUBLIC kenlm utils fst) target_link_libraries(decoder PUBLIC kenlm utils fst)
\ No newline at end of file
...@@ -15,7 +15,7 @@ ...@@ -15,7 +15,7 @@
#include "base/common.h" #include "base/common.h"
#include "decoder/ctc_decoders/path_trie.h" #include "decoder/ctc_decoders/path_trie.h"
#include "decoder/ctc_decoders/scorer.h" #include "decoder/ctc_decoders/scorer.h"
#include "nnet/decodable-itf.h" #include "kaldi/decoder/decodable-itf.h"
#include "util/parse-options.h" #include "util/parse-options.h"
#pragma once #pragma once
......
#include "decoder/ctc_tlg_decoder.h"
namespace ppspeech {
TLGDecoder::TLGDecoder(TLGDecoderOptions opts) {
fst_.reset(fst::Fst<fst::StdArc>::Read(opts.fst_path));
CHECK(fst_ != nullptr);
word_symbol_table_.reset(fst::SymbolTable::ReadText(opts.word_symbol_table));
decoder_.reset(new kaldi::LatticeFasterOnlineDecoder(*fst_, opts.opts));
decoder_->InitDecoding();
}
void TLGDecoder::InitDecoder() {
decoder_->InitDecoding();
}
void TLGDecoder::AdvanceDecode(const std::shared_ptr<kaldi::DecodableInterface>& decodable) {
while (1) {
AdvanceDecoding(decodable.get());
if (decodable->IsLastFrame(num_frame_decoded_)) break;
}
}
void TLGDecoder::AdvanceDecoding(kaldi::DecodableInterface* decodable) {
// skip blank frame?
decoder_->AdvanceDecoding(decodable, 1);
num_frame_decoded_++;
}
void TLGDecoder::Reset() {
decoder_->InitDecoding();
return;
}
std::string TLGDecoder::GetFinalBestPath() {
decoder_->FinalizeDecoding();
kaldi::Lattice lat;
kaldi::LatticeWeight weight;
std::vector<int> alignment;
std::vector<int> words_id;
decoder_->GetBestPath(&lat, true);
fst::GetLinearSymbolSequence(lat, &alignment, &words_id, &weight);
std::string words;
for (int32 idx = 0; idx < words_id.size(); ++idx) {
std::string word = word_symbol_table_->Find(words_id[idx]);
words += word;
}
return words;
}
}
\ No newline at end of file
#pragma once
#include "kaldi/decoder/lattice-faster-online-decoder.h"
#include "kaldi/decoder/decodable-itf.h"
#include "util/parse-options.h"
#include "base/basic_types.h"
namespace ppspeech {
struct TLGDecoderOptions {
kaldi::LatticeFasterDecoderConfig opts;
// todo remove later, add into decode resource
std::string word_symbol_table;
std::string fst_path;
TLGDecoderOptions()
: word_symbol_table(""),
fst_path("") {}
};
class TLGDecoder {
public:
explicit TLGDecoder(TLGDecoderOptions opts);
void InitDecoder();
void Decode();
std::string GetBestPath();
std::vector<std::pair<double, std::string>> GetNBestPath();
std::string GetFinalBestPath();
int NumFrameDecoded();
int DecodeLikelihoods(const std::vector<std::vector<BaseFloat>>& probs,
std::vector<std::string>& nbest_words);
void AdvanceDecode(
const std::shared_ptr<kaldi::DecodableInterface>& decodable);
void Reset();
private:
void AdvanceDecoding(kaldi::DecodableInterface* decodable);
std::shared_ptr<kaldi::LatticeFasterOnlineDecoder> decoder_;
std::shared_ptr<fst::Fst<fst::StdArc>> fst_;
std::shared_ptr<fst::SymbolTable> word_symbol_table_;
int32 num_frame_decoded_;
};
} // namespace ppspeech
\ No newline at end of file
...@@ -4,3 +4,6 @@ add_subdirectory(base) ...@@ -4,3 +4,6 @@ add_subdirectory(base)
add_subdirectory(util) add_subdirectory(util)
add_subdirectory(feat) add_subdirectory(feat)
add_subdirectory(matrix) add_subdirectory(matrix)
add_subdirectory(lat)
add_subdirectory(fstext)
add_subdirectory(decoder)
add_library(kaldi-decoder
lattice-faster-decoder.cc
lattice-faster-online-decoder.cc
)
target_link_libraries(kaldi-decoder PUBLIC kaldi-lat)
...@@ -121,7 +121,7 @@ class DecodableInterface { ...@@ -121,7 +121,7 @@ class DecodableInterface {
/// decoding-from-matrix setting where we want to allow the last delta or /// decoding-from-matrix setting where we want to allow the last delta or
/// LDA /// LDA
/// features to be flushed out for compatibility with the baseline setup. /// features to be flushed out for compatibility with the baseline setup.
virtual bool IsLastFrame(int32 frame) const = 0; virtual bool IsLastFrame(int32 frame) = 0;
/// The call NumFramesReady() will return the number of frames currently /// The call NumFramesReady() will return the number of frames currently
/// available /// available
......
...@@ -1007,14 +1007,10 @@ template class LatticeFasterDecoderTpl<fst::Fst<fst::StdArc>, decoder::StdToken> ...@@ -1007,14 +1007,10 @@ template class LatticeFasterDecoderTpl<fst::Fst<fst::StdArc>, decoder::StdToken>
template class LatticeFasterDecoderTpl<fst::VectorFst<fst::StdArc>, decoder::StdToken >; template class LatticeFasterDecoderTpl<fst::VectorFst<fst::StdArc>, decoder::StdToken >;
template class LatticeFasterDecoderTpl<fst::ConstFst<fst::StdArc>, decoder::StdToken >; template class LatticeFasterDecoderTpl<fst::ConstFst<fst::StdArc>, decoder::StdToken >;
template class LatticeFasterDecoderTpl<fst::ConstGrammarFst, decoder::StdToken>;
template class LatticeFasterDecoderTpl<fst::VectorGrammarFst, decoder::StdToken>;
template class LatticeFasterDecoderTpl<fst::Fst<fst::StdArc> , decoder::BackpointerToken>; template class LatticeFasterDecoderTpl<fst::Fst<fst::StdArc> , decoder::BackpointerToken>;
template class LatticeFasterDecoderTpl<fst::VectorFst<fst::StdArc>, decoder::BackpointerToken >; template class LatticeFasterDecoderTpl<fst::VectorFst<fst::StdArc>, decoder::BackpointerToken >;
template class LatticeFasterDecoderTpl<fst::ConstFst<fst::StdArc>, decoder::BackpointerToken >; template class LatticeFasterDecoderTpl<fst::ConstFst<fst::StdArc>, decoder::BackpointerToken >;
template class LatticeFasterDecoderTpl<fst::ConstGrammarFst, decoder::BackpointerToken>;
template class LatticeFasterDecoderTpl<fst::VectorGrammarFst, decoder::BackpointerToken>;
} // end namespace kaldi. } // end namespace kaldi.
...@@ -23,11 +23,10 @@ ...@@ -23,11 +23,10 @@
#ifndef KALDI_DECODER_LATTICE_FASTER_DECODER_H_ #ifndef KALDI_DECODER_LATTICE_FASTER_DECODER_H_
#define KALDI_DECODER_LATTICE_FASTER_DECODER_H_ #define KALDI_DECODER_LATTICE_FASTER_DECODER_H_
#include "decoder/grammar-fst.h"
#include "fst/fstlib.h" #include "fst/fstlib.h"
#include "fst/memory.h" #include "fst/memory.h"
#include "fstext/fstext-lib.h" #include "fstext/fstext-lib.h"
#include "itf/decodable-itf.h" #include "decoder/decodable-itf.h"
#include "lat/determinize-lattice-pruned.h" #include "lat/determinize-lattice-pruned.h"
#include "lat/kaldi-lattice.h" #include "lat/kaldi-lattice.h"
#include "util/hash-list.h" #include "util/hash-list.h"
......
...@@ -278,8 +278,8 @@ bool LatticeFasterOnlineDecoderTpl<FST>::GetRawLatticePruned( ...@@ -278,8 +278,8 @@ bool LatticeFasterOnlineDecoderTpl<FST>::GetRawLatticePruned(
template class LatticeFasterOnlineDecoderTpl<fst::Fst<fst::StdArc> >; template class LatticeFasterOnlineDecoderTpl<fst::Fst<fst::StdArc> >;
template class LatticeFasterOnlineDecoderTpl<fst::VectorFst<fst::StdArc> >; template class LatticeFasterOnlineDecoderTpl<fst::VectorFst<fst::StdArc> >;
template class LatticeFasterOnlineDecoderTpl<fst::ConstFst<fst::StdArc> >; template class LatticeFasterOnlineDecoderTpl<fst::ConstFst<fst::StdArc> >;
template class LatticeFasterOnlineDecoderTpl<fst::ConstGrammarFst >; //template class LatticeFasterOnlineDecoderTpl<fst::ConstGrammarFst >;
template class LatticeFasterOnlineDecoderTpl<fst::VectorGrammarFst >; //template class LatticeFasterOnlineDecoderTpl<fst::VectorGrammarFst >;
} // end namespace kaldi. } // end namespace kaldi.
...@@ -30,7 +30,7 @@ ...@@ -30,7 +30,7 @@
#include "util/stl-utils.h" #include "util/stl-utils.h"
#include "util/hash-list.h" #include "util/hash-list.h"
#include "fst/fstlib.h" #include "fst/fstlib.h"
#include "itf/decodable-itf.h" #include "decoder/decodable-itf.h"
#include "fstext/fstext-lib.h" #include "fstext/fstext-lib.h"
#include "lat/determinize-lattice-pruned.h" #include "lat/determinize-lattice-pruned.h"
#include "lat/kaldi-lattice.h" #include "lat/kaldi-lattice.h"
......
add_library(kaldi-fstext
kaldi-fst-io.cc
)
target_link_libraries(kaldi-fstext PUBLIC kaldi-util)
此差异已折叠。
// fstext/determinize-lattice.h
// Copyright 2009-2011 Microsoft Corporation
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#ifndef KALDI_FSTEXT_DETERMINIZE_LATTICE_H_
#define KALDI_FSTEXT_DETERMINIZE_LATTICE_H_
#include <fst/fst-decl.h>
#include <fst/fstlib.h>
#include <algorithm>
#include <map>
#include <set>
#include <vector>
#include "fstext/lattice-weight.h"
namespace fst {
/// \addtogroup fst_extensions
/// @{
// For example of usage, see test-determinize-lattice.cc
/*
DeterminizeLattice implements a special form of determinization
with epsilon removal, optimized for a phase of lattice generation.
Its input is an FST with weight-type BaseWeightType (usually a pair of
floats, with a lexicographical type of order, such as
LatticeWeightTpl<float>). Typically this would be a state-level lattice, with
input symbols equal to words, and output-symbols equal to p.d.f's (so like
the inverse of HCLG). Imagine representing this as an acceptor of type
CompactLatticeWeightTpl<float>, in which the input/output symbols are words,
and the weights contain the original weights together with strings (with zero
or one symbol in them) containing the original output labels (the p.d.f.'s).
We determinize this using acceptor determinization with epsilon removal.
Remember (from lattice-weight.h) that CompactLatticeWeightTpl has a special
kind of semiring where we always take the string corresponding to the best
cost (of type BaseWeightType), and discard the other. This corresponds to
taking the best output-label sequence (of p.d.f.'s) for each input-label
sequence (of words). We couldn't use the Gallic weight for this, or it would
die as soon as it detected that the input FST was non-functional. In our
case, any acyclic FST (and many cyclic ones) can be determinized. We assume
that there is a function Compare(const BaseWeightType &a, const
BaseWeightType &b) that returns (-1, 0, 1) according to whether (a < b, a ==
b, a > b) in the total order on the BaseWeightType... this information should
be the same as NaturalLess would give, but it's more efficient to do it this
way. You can define this for things like TropicalWeight if you need to
instantiate this class for that weight type.
We implement this determinization in a special way to make it efficient for
the types of FSTs that we will apply it to. One issue is that if we
explicitly represent the strings (in CompactLatticeWeightTpl) as vectors of
type vector<IntType>, the algorithm takes time quadratic in the length of
words (in states), because propagating each arc involves copying a whole
vector (of integers representing p.d.f.'s). Instead we use a hash structure
where each string is a pointer (Entry*), and uses a hash from (Entry*,
IntType), to the successor string (and a way to get the latest IntType and
the ancestor Entry*). [this is the class LatticeStringRepository].
Another issue is that rather than representing a determinized-state as a
collection of (state, weight), we represent it in a couple of reduced forms.
Suppose a determinized-state is a collection of (state, weight) pairs; call
this the "canonical representation". Note: these collections are always
normalized to remove any common weight and string part. Define end-states as
the subset of states that have an arc out of them with a label on, or are
final. If we represent a determinized-state a the set of just its
(end-state, weight) pairs, this will be a valid and more compact
representation, and will lead to a smaller set of determinized states (like
early minimization). Call this collection of (end-state, weight) pairs the
"minimal representation". As a mechanism to reduce compute, we can also
consider another representation. In the determinization algorithm, we start
off with a set of (begin-state, weight) pairs (where the "begin-states" are
initial or have a label on the transition into them), and the "canonical
representation" consists of the epsilon-closure of this set (i.e. follow
epsilons). Call this set of (begin-state, weight) pairs, appropriately
normalized, the "initial representation". If two initial representations are
the same, the "canonical representation" and hence the "minimal
representation" will be the same. We can use this to reduce compute. Note
that if two initial representations are different, this does not preclude the
other representations from being the same.
*/
struct DeterminizeLatticeOptions {
float delta; // A small offset used to measure equality of weights.
int max_mem; // If >0, determinization will fail and return false
// when the algorithm's (approximate) memory consumption crosses this
// threshold.
int max_loop; // If >0, can be used to detect non-determinizable input
// (a case that wouldn't be caught by max_mem).
DeterminizeLatticeOptions() : delta(kDelta), max_mem(-1), max_loop(-1) {}
};
/**
This function implements the normal version of DeterminizeLattice, in which
the output strings are represented using sequences of arcs, where all but
the first one has an epsilon on the input side. The debug_ptr argument is
an optional pointer to a bool that, if it becomes true while the algorithm
is executing, the algorithm will print a traceback and terminate (used in
fstdeterminizestar.cc debug non-terminating determinization). More
efficient if ifst is arc-sorted on input label. If the number of arcs gets
more than max_states, it will throw std::runtime_error (otherwise this code
does not use exceptions). This is mainly useful for debug. */
template <class Weight, class IntType>
bool DeterminizeLattice(
const Fst<ArcTpl<Weight> > &ifst, MutableFst<ArcTpl<Weight> > *ofst,
DeterminizeLatticeOptions opts = DeterminizeLatticeOptions(),
bool *debug_ptr = NULL);
/* This is a version of DeterminizeLattice with a slightly more "natural"
output format, where the output sequences are encoded using the
CompactLatticeArcTpl template (i.e. the sequences of output symbols are
represented directly as strings) More efficient if ifst is arc-sorted on
input label. If the #arcs gets more than max_arcs, it will throw
std::runtime_error (otherwise this code does not use exceptions). This is
mainly useful for debug.
*/
template <class Weight, class IntType>
bool DeterminizeLattice(
const Fst<ArcTpl<Weight> > &ifst,
MutableFst<ArcTpl<CompactLatticeWeightTpl<Weight, IntType> > > *ofst,
DeterminizeLatticeOptions opts = DeterminizeLatticeOptions(),
bool *debug_ptr = NULL);
/// @} end "addtogroup fst_extensions"
} // end namespace fst
#include "fstext/determinize-lattice-inl.h"
#endif // KALDI_FSTEXT_DETERMINIZE_LATTICE_H_
此差异已折叠。
// fstext/determinize-star.h
// Copyright 2009-2011 Microsoft Corporation
// 2014 Guoguo Chen
// 2015 Hainan Xu
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#ifndef KALDI_FSTEXT_DETERMINIZE_STAR_H_
#define KALDI_FSTEXT_DETERMINIZE_STAR_H_
#include <fst/fst-decl.h>
#include <fst/fstlib.h>
#include <algorithm>
#include <map>
#include <set>
#include <stdexcept> // this algorithm uses exceptions
#include <vector>
namespace fst {
/// \addtogroup fst_extensions
/// @{
// For example of usage, see test-determinize-star.cc
/*
DeterminizeStar implements determinization with epsilon removal, which we
distinguish with a star.
We define a determinized* FST as one in which no state has more than one
transition with the same input-label. Epsilon input labels are not allowed
except starting from states that have exactly one arc exiting them (and are
not final). [In the normal definition of determinized, epsilon-input labels
are not allowed at all, whereas in Mohri's definition, epsilons are treated
as ordinary symbols]. The determinized* definition is intended to simulate
the effect of allowing strings of output symbols at each state.
The algorithm implemented here takes an Fst<Arc>, and a pointer to a
MutableFst<Arc> where it puts its output. The weight type is assumed to be a
float-weight. It does epsilon removal and determinization.
This algorithm may fail if the input has epsilon cycles under
certain circumstances (i.e. the semiring is non-idempotent, e.g. the log
semiring, or there are negative cost epsilon cycles).
This implementation is much less fancy than the one in fst/determinize.h, and
does not have an "on-demand" version.
The algorithm is a fairly normal determinization algorithm. We keep in
memory the subsets of states, together with their leftover strings and their
weights. The only difference is we detect input epsilon transitions and
treat them "specially".
*/
// This algorithm will be slightly faster if you sort the input fst on input
// label.
/**
This function implements the normal version of DeterminizeStar, in which the
output strings are represented using sequences of arcs, where all but the
first one has an epsilon on the input side. The debug_ptr argument is an
optional pointer to a bool that, if it becomes true while the algorithm is
executing, the algorithm will print a traceback and terminate (used in
fstdeterminizestar.cc debug non-terminating determinization).
If max_states is positive, it will stop determinization and throw an
exception as soon as the max-states is reached. This can be useful in test.
If allow_partial is true, the algorithm will output partial results when the
specified max_states is reached (when larger than zero), instead of throwing
out an error.
Caution, the return status is un-intuitive: this function will return false
if determinization completed normally, and true if it was stopped early by
reaching the 'max-states' limit, and a partial FST was generated.
*/
template <class F>
bool DeterminizeStar(F &ifst, MutableFst<typename F::Arc> *ofst, // NOLINT
float delta = kDelta, bool *debug_ptr = NULL,
int max_states = -1, bool allow_partial = false);
/* This is a version of DeterminizeStar with a slightly more "natural" output
format, where the output sequences are encoded using the GallicArc (i.e. the
output symbols are strings. If max_states is positive, it will stop
determinization and throw an exception as soon as the max-states is reached.
This can be useful in test. If allow_partial is true, the algorithm will
output partial results when the specified max_states is reached (when larger
than zero), instead of throwing out an error.
Caution, the return status is un-intuitive: this function will return false
if determinization completed normally, and true if it was stopped early by
reaching the 'max-states' limit, and a partial FST was generated.
*/
template <class F>
bool DeterminizeStar(F &ifst, // NOLINT
MutableFst<GallicArc<typename F::Arc> > *ofst,
float delta = kDelta, bool *debug_ptr = NULL,
int max_states = -1, bool allow_partial = false);
/// @} end "addtogroup fst_extensions"
} // end namespace fst
#include "fstext/determinize-star-inl.h"
#endif // KALDI_FSTEXT_DETERMINIZE_STAR_H_
// fstext/fstext-lib.h
// Copyright 2009-2012 Microsoft Corporation Johns Hopkins University (author:
// Daniel Povey)
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#ifndef KALDI_FSTEXT_FSTEXT_LIB_H_
#define KALDI_FSTEXT_FSTEXT_LIB_H_
#include "fst/fstlib.h"
#include "fstext/determinize-lattice.h"
#include "fstext/determinize-star.h"
#include "fstext/fstext-utils.h"
#include "fstext/kaldi-fst-io.h"
#include "fstext/lattice-utils.h"
#include "fstext/lattice-weight.h"
#include "fstext/pre-determinize.h"
#include "fstext/table-matcher.h"
#endif // KALDI_FSTEXT_FSTEXT_LIB_H_
此差异已折叠。
// fstext/fstext-utils.h
// Copyright 2009-2011 Microsoft Corporation
// 2012-2013 Johns Hopkins University (Author: Daniel Povey)
// 2013 Guoguo Chen
// 2014 Telepoint Global Hosting Service, LLC. (Author: David
// Snyder)
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#ifndef KALDI_FSTEXT_FSTEXT_UTILS_H_
#define KALDI_FSTEXT_FSTEXT_UTILS_H_
#include <fst/fst-decl.h>
#include <fst/fstlib.h>
#include <algorithm>
#include <map>
#include <set>
#include <vector>
#include "fstext/determinize-star.h"
#include "fstext/remove-eps-local.h"
#include "base/kaldi-common.h" // for error reporting macros.
#include "util/text-utils.h" // for SplitStringToVector
#include "fst/script/print-impl.h"
namespace fst {
/// Returns the highest numbered output symbol id of the FST (or zero
/// for an empty FST.
template <class Arc>
typename Arc::Label HighestNumberedOutputSymbol(const Fst<Arc> &fst);
/// Returns the highest numbered input symbol id of the FST (or zero
/// for an empty FST.
template <class Arc>
typename Arc::Label HighestNumberedInputSymbol(const Fst<Arc> &fst);
/// Returns the total number of arcs in an FST.
template <class Arc>
typename Arc::StateId NumArcs(const ExpandedFst<Arc> &fst);
/// GetInputSymbols gets the list of symbols on the input of fst
/// (including epsilon, if include_eps == true), as a sorted, unique
/// list.
template <class Arc, class I>
void GetInputSymbols(const Fst<Arc> &fst, bool include_eps,
std::vector<I> *symbols);
/// GetOutputSymbols gets the list of symbols on the output of fst
/// (including epsilon, if include_eps == true)
template <class Arc, class I>
void GetOutputSymbols(const Fst<Arc> &fst, bool include_eps,
std::vector<I> *symbols);
/// ClearSymbols sets all the symbols on the input and/or
/// output side of the FST to zero, as specified.
/// It does not alter the symbol tables.
template <class Arc>
void ClearSymbols(bool clear_input, bool clear_output, MutableFst<Arc> *fst);
template <class I>
void GetSymbols(const SymbolTable &symtab, bool include_eps,
std::vector<I> *syms_out);
inline void DeterminizeStarInLog(VectorFst<StdArc> *fst, float delta = kDelta,
bool *debug_ptr = NULL, int max_states = -1);
// e.g. of using this function: PushInLog<REWEIGHT_TO_INITIAL>(fst,
// kPushWeights|kPushLabels);
template <ReweightType rtype> // == REWEIGHT_TO_{INITIAL, FINAL}
void PushInLog(VectorFst<StdArc> *fst, uint32 ptype, float delta = kDelta) {
// PushInLog pushes the FST
// and returns a new pushed FST (labels and weights pushed to the left).
VectorFst<LogArc> *fst_log =
new VectorFst<LogArc>; // Want to determinize in log semiring.
Cast(*fst, fst_log);
VectorFst<StdArc> tmp;
*fst = tmp; // free up memory.
VectorFst<LogArc> *fst_pushed_log = new VectorFst<LogArc>;
Push<LogArc, rtype>(*fst_log, fst_pushed_log, ptype, delta);
Cast(*fst_pushed_log, fst);
delete fst_log;
delete fst_pushed_log;
}
// Minimizes after encoding; applicable to all FSTs. It is like what you get
// from the Minimize() function, except it will not push the weights, or the
// symbols. This is better for our recipes, as we avoid ever pushing the
// weights. However, it will only minimize optimally if your graphs are such
// that the symbols are as far to the left as they can go, and the weights
// in combinable paths are the same... hard to formalize this, but it's
// something that is satisified by our normal FSTs.
template <class Arc>
void MinimizeEncoded(VectorFst<Arc> *fst, float delta = kDelta) {
Map(fst, QuantizeMapper<Arc>(delta));
EncodeMapper<Arc> encoder(kEncodeLabels | kEncodeWeights, ENCODE);
Encode(fst, &encoder);
internal::AcceptorMinimize(fst);
Decode(fst, encoder);
}
/// GetLinearSymbolSequence gets the symbol sequence from a linear FST.
/// If the FST is not just a linear sequence, it returns false. If it is
/// a linear sequence (including the empty FST), it returns true. In this
/// case it outputs the symbol
/// sequences as "isymbols_out" and "osymbols_out" (removing epsilons), and
/// the total weight as "tot_weight". The total weight will be Weight::Zero()
/// if the FST is empty. If any of the output pointers are NULL, it does not
/// create that output.
template <class Arc, class I>
bool GetLinearSymbolSequence(const Fst<Arc> &fst, std::vector<I> *isymbols_out,
std::vector<I> *osymbols_out,
typename Arc::Weight *tot_weight_out);
/// This function converts an FST with a special structure, which is
/// output by the OpenFst functions ShortestPath and RandGen, and converts
/// them into a std::vector of separate FSTs. This special structure is that
/// the only state that has more than one (arcs-out or final-prob) is the
/// start state. fsts_out is resized to the appropriate size.
template <class Arc>
void ConvertNbestToVector(const Fst<Arc> &fst,
std::vector<VectorFst<Arc> > *fsts_out);
/// Takes the n-shortest-paths (using ShortestPath), but outputs
/// the result as a vector of up to n fsts. This function will
/// size the "fsts_out" vector to however many paths it got
/// (which will not exceed n). n must be >= 1.
template <class Arc>
void NbestAsFsts(const Fst<Arc> &fst, size_t n,
std::vector<VectorFst<Arc> > *fsts_out);
/// Creates unweighted linear acceptor from symbol sequence.
template <class Arc, class I>
void MakeLinearAcceptor(const std::vector<I> &labels, MutableFst<Arc> *ofst);
/// Creates an unweighted acceptor with a linear structure, with alternatives
/// at each position. Epsilon is treated like a normal symbol here.
/// Each position in "labels" must have at least one alternative.
template <class Arc, class I>
void MakeLinearAcceptorWithAlternatives(
const std::vector<std::vector<I> > &labels, MutableFst<Arc> *ofst);
/// Does PreDeterminize and DeterminizeStar and then removes the disambiguation
/// symbols. This is a form of determinization that will never blow up. Note
/// that ifst is non-const and can be considered to be destroyed by this
/// operation.
/// Does not do epsilon removal (RemoveEpsLocal)-- this is so it's safe to cast
/// to log and do this, and maintain equivalence in tropical.
template <class Arc>
void SafeDeterminizeWrapper(MutableFst<Arc> *ifst, MutableFst<Arc> *ofst,
float delta = kDelta);
/// SafeDeterminizeMinimizeWapper is as SafeDeterminizeWrapper except that it
/// also minimizes (encoded minimization, which is safe). This algorithm will
/// destroy "ifst".
template <class Arc>
void SafeDeterminizeMinimizeWrapper(MutableFst<Arc> *ifst, VectorFst<Arc> *ofst,
float delta = kDelta);
/// SafeDeterminizeMinimizeWapperInLog is as SafeDeterminizeMinimizeWrapper
/// except it first casts tothe log semiring.
void SafeDeterminizeMinimizeWrapperInLog(VectorFst<StdArc> *ifst,
VectorFst<StdArc> *ofst,
float delta = kDelta);
/// RemoveSomeInputSymbols removes any symbol that appears in "to_remove", from
/// the input side of the FST, replacing them with epsilon.
template <class Arc, class I>
void RemoveSomeInputSymbols(const std::vector<I> &to_remove,
MutableFst<Arc> *fst);
// MapInputSymbols will replace any input symbol i that is between 0 and
// symbol_map.size()-1, with symbol_map[i]. It removes the input symbol
// table of the FST.
template <class Arc, class I>
void MapInputSymbols(const std::vector<I> &symbol_map, MutableFst<Arc> *fst);
template <class Arc>
void RemoveWeights(MutableFst<Arc> *fst);
/// Returns true if and only if the FST is such that the input symbols
/// on arcs entering any given state all have the same value.
/// if "start_is_epsilon", treat start-state as an epsilon input arc
/// [i.e. ensure only epsilon can enter start-state].
template <class Arc>
bool PrecedingInputSymbolsAreSame(bool start_is_epsilon, const Fst<Arc> &fst);
/// This is as PrecedingInputSymbolsAreSame, but with a functor f that maps
/// labels to classes. The function tests whether the symbols preceding any
/// given state are in the same class. Formally, f is of a type F that has an
/// operator of type F::Result F::operator() (F::Arg a) const; where F::Result
/// is an integer type and F::Arc can be constructed from Arc::Label. this must
/// apply to valid labels and also to kNoLabel (so we can have a marker for the
/// invalid labels.
template <class Arc, class F>
bool PrecedingInputSymbolsAreSameClass(bool start_is_epsilon,
const Fst<Arc> &fst, const F &f);
/// Returns true if and only if the FST is such that the input symbols
/// on arcs exiting any given state all have the same value.
/// If end_is_epsilon, treat end-state as an epsilon output arc [i.e. ensure
/// end-states cannot have non-epsilon output transitions.]
template <class Arc>
bool FollowingInputSymbolsAreSame(bool end_is_epsilon, const Fst<Arc> &fst);
template <class Arc, class F>
bool FollowingInputSymbolsAreSameClass(bool end_is_epsilon, const Fst<Arc> &fst,
const F &f);
/// MakePrecedingInputSymbolsSame ensures that all arcs entering any given fst
/// state have the same input symbol. It does this by detecting states
/// that have differing input symbols going in, and inserting, for each of
/// the preceding arcs with non-epsilon input symbol, a new dummy state that
/// has an epsilon link to the fst state.
/// If "start_is_epsilon", ensure that start-state can have only epsilon-links
/// into it.
template <class Arc>
void MakePrecedingInputSymbolsSame(bool start_is_epsilon, MutableFst<Arc> *fst);
/// As MakePrecedingInputSymbolsSame, but takes a functor object that maps
/// labels to classes.
template <class Arc, class F>
void MakePrecedingInputSymbolsSameClass(bool start_is_epsilon,
MutableFst<Arc> *fst, const F &f);
/// MakeFollowingInputSymbolsSame ensures that all arcs exiting any given fst
/// state have the same input symbol. It does this by detecting states that
/// have differing input symbols on arcs that exit it, and inserting, for each
/// of the following arcs with non-epsilon input symbol, a new dummy state that
/// has an input-epsilon link from the fst state. The output symbol and weight
/// stay on the link to the dummy state (in order to keep the FST
/// output-deterministic and stochastic, if it already was). If end_is_epsilon,
/// treat "being a final-state" like having an epsilon output link.
template <class Arc>
void MakeFollowingInputSymbolsSame(bool end_is_epsilon, MutableFst<Arc> *fst);
/// As MakeFollowingInputSymbolsSame, but takes a functor object that maps
/// labels to classes.
template <class Arc, class F>
void MakeFollowingInputSymbolsSameClass(bool end_is_epsilon,
MutableFst<Arc> *fst, const F &f);
/// MakeLoopFst creates an FST that has a state that is both initial and
/// final (weight == Weight::One()), and for each non-NULL pointer fsts[i],
/// it has an arc out whose output-symbol is i and which goes to a
/// sub-graph whose input language is equivalent to fsts[i], where the
/// final-state becomes a transition to the loop-state. Each fst in "fsts"
/// should be an acceptor. The fst MakeLoopFst returns is output-deterministic,
/// but not output-epsilon free necessarily, and arcs are sorted on output
/// label. Note: if some of the pointers in the input vector "fsts" have the
/// same value, "MakeLoopFst" uses this to speed up the computation.
/// Formally: suppose I is the set of indexes i such that fsts[i] != NULL.
/// Let L[i] be the language that the acceptor fsts[i] accepts.
/// Let the language K be the set of input-output pairs i:l such
/// that i in I and l in L[i]. Then the FST returned by MakeLoopFst
/// accepts the language K*, where * is the Kleene closure (CLOSURE_STAR)
/// of K.
/// We could have implemented this via a combination of "project",
/// "concat", "union" and "closure". But that FST would have been
/// less well optimized and would have a lot of final-states.
template <class Arc>
VectorFst<Arc> *MakeLoopFst(const std::vector<const ExpandedFst<Arc> *> &fsts);
/// ApplyProbabilityScale is applicable to FSTs in the log or tropical semiring.
/// It multiplies the arc and final weights by "scale" [this is not the Mul
/// operation of the semiring, it's actual multiplication, which is equivalent
/// to taking a power in the semiring].
template <class Arc>
void ApplyProbabilityScale(float scale, MutableFst<Arc> *fst);
/// EqualAlign is similar to RandGen, but it generates a sequence with exactly
/// "length" input symbols. It returns true on success, false on failure
/// (failure is partly random but should never happen in practice for normal
/// speech models.) It generates a random path through the input FST, finds out
/// which subset of the states it visits along the way have self-loops with
/// inupt symbols on them, and outputs a path with exactly enough self-loops to
/// have the requested number of input symbols. Note that EqualAlign does not
/// use the probabilities on the FST. It just uses equal probabilities in the
/// first stage of selection (since the output will anyway not be a truly random
/// sample from the FST). The input fst "ifst" must be connected or this may
/// enter an infinite loop.
template <class Arc>
bool EqualAlign(const Fst<Arc> &ifst, typename Arc::StateId length,
int rand_seed, MutableFst<Arc> *ofst, int num_retries = 10);
// RemoveUselessArcs removes arcs such that there is no input symbol
// sequence for which the best path through the FST would contain
// those arcs [for these purposes, epsilon is not treated as a real symbol].
// This is mainly geared towards decoding-graph FSTs which may contain
// transitions that have less likely words on them that would never be
// taken. We do not claim that this algorithm removes all such arcs;
// it just does the best job it can.
// Only works for tropical (not log) semiring as it uses
// NaturalLess.
template <class Arc>
void RemoveUselessArcs(MutableFst<Arc> *fst);
// PhiCompose is a version of composition where
// the right hand FST (fst2) is treated as a backoff
// LM, with the phi symbol (e.g. #0) treated as a
// "failure transition", only taken when we don't
// have a match for the requested symbol.
template <class Arc>
void PhiCompose(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
typename Arc::Label phi_label, MutableFst<Arc> *fst);
// PropagateFinal propagates final-probs through
// "phi" transitions (note that here, phi_label may
// be epsilon if you want). If you have a backoff LM
// with special symbols ("phi") on the backoff arcs
// instead of epsilon, you may use PhiCompose to compose
// with it, but this won't do the right thing w.r.t.
// final probabilities. You should first call PropagateFinal
// on the FST with phi's i it (fst2 in PhiCompose above),
// to fix this. If a state does not have a final-prob,
// but has a phi transition, it makes the state's final-prob
// (phi-prob * final-prob-of-dest-state), and does this
// recursively i.e. follows phi transitions on the dest state
// first. It behaves as if there were a super-final state
// with a special symbol leading to it, from each currently
// final state. Note that this may not behave as desired
// if there are epsilons in your FST; it might be better
// to remove those before calling this function.
template <class Arc>
void PropagateFinal(typename Arc::Label phi_label, MutableFst<Arc> *fst);
// PhiCompose is a version of composition where
// the right hand FST (fst2) has speciall "rho transitions"
// which are taken whenever no normal transition matches; these
// transitions will be rewritten with whatever symbol was on
// the first FST.
template <class Arc>
void RhoCompose(const Fst<Arc> &fst1, const Fst<Arc> &fst2,
typename Arc::Label rho_label, MutableFst<Arc> *fst);
/** This function returns true if, in the semiring of the FST, the sum (within
the semiring) of all the arcs out of each state in the FST is one, to within
delta. After MakeStochasticFst, this should be true (for a connected FST).
@param fst [in] the FST that we are testing.
@param delta [in] the tolerance to within which we test equality to 1.
@param min_sum [out] if non, NULL, contents will be set to the minimum sum
of weights.
@param max_sum [out] if non, NULL, contents will be set to the maximum sum
of weights.
@return Returns true if the FST is stochastic, and false otherwise.
*/
template <class Arc>
bool IsStochasticFst(const Fst<Arc> &fst,
float delta = kDelta, // kDelta = 1.0/1024.0 by default.
typename Arc::Weight *min_sum = NULL,
typename Arc::Weight *max_sum = NULL);
// IsStochasticFstInLog makes sure it's stochastic after casting to log.
inline bool IsStochasticFstInLog(
const Fst<StdArc> &fst,
float delta = kDelta, // kDelta = 1.0/1024.0 by default.
StdArc::Weight *min_sum = NULL, StdArc::Weight *max_sum = NULL);
} // end namespace fst
#include "fstext/fstext-utils-inl.h"
#endif // KALDI_FSTEXT_FSTEXT_UTILS_H_
此差异已折叠。
// fstext/kaldi-fst-io.cc
// Copyright 2009-2011 Microsoft Corporation
// 2012-2015 Johns Hopkins University (Author: Daniel Povey)
// 2013 Guoguo Chen
// See ../../COPYING for clarification regarding multiple authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// THIS CODE IS PROVIDED *AS IS* BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, EITHER EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION ANY IMPLIED
// WARRANTIES OR CONDITIONS OF TITLE, FITNESS FOR A PARTICULAR PURPOSE,
// MERCHANTABLITY OR NON-INFRINGEMENT.
// See the Apache 2 License for the specific language governing permissions and
// limitations under the License.
#include "fstext/kaldi-fst-io.h"
#include <string>
#include "base/kaldi-error.h"
#include "base/kaldi-math.h"
#include "util/kaldi-io.h"
namespace fst {
VectorFst<StdArc> *ReadFstKaldi(std::string rxfilename) {
if (rxfilename == "") rxfilename = "-"; // interpret "" as stdin,
// for compatibility with OpenFst conventions.
kaldi::Input ki(rxfilename);
fst::FstHeader hdr;
if (!hdr.Read(ki.Stream(), rxfilename))
KALDI_ERR << "Reading FST: error reading FST header from "
<< kaldi::PrintableRxfilename(rxfilename);
FstReadOptions ropts("<unspecified>", &hdr);
VectorFst<StdArc> *fst = VectorFst<StdArc>::Read(ki.Stream(), ropts);
if (!fst)
KALDI_ERR << "Could not read fst from "
<< kaldi::PrintableRxfilename(rxfilename);
return fst;
}
// Register const fst to load it automatically. Other types like
// olabel_lookahead or ngram or compact_fst should be registered
// through OpenFst registration API.
static fst::FstRegisterer<VectorFst<StdArc>> VectorFst_StdArc_registerer;
static fst::FstRegisterer<ConstFst<StdArc>> ConstFst_StdArc_registerer;
Fst<StdArc> *ReadFstKaldiGeneric(std::string rxfilename, bool throw_on_err) {
if (rxfilename == "") rxfilename = "-"; // interpret "" as stdin,
// for compatibility with OpenFst conventions.
kaldi::Input ki(rxfilename);
fst::FstHeader hdr;
// Read FstHeader which contains the type of FST
if (!hdr.Read(ki.Stream(), rxfilename)) {
if (throw_on_err) {
KALDI_ERR << "Reading FST: error reading FST header from "
<< kaldi::PrintableRxfilename(rxfilename);
} else {
KALDI_WARN << "We fail to read FST header from "
<< kaldi::PrintableRxfilename(rxfilename)
<< ". A NULL pointer is returned.";
return NULL;
}
}
// Check the type of Arc
if (hdr.ArcType() != fst::StdArc::Type()) {
if (throw_on_err) {
KALDI_ERR << "FST with arc type " << hdr.ArcType()
<< " is not supported.";
} else {
KALDI_WARN << "Fst with arc type" << hdr.ArcType()
<< " is not supported. A NULL pointer is returned.";
return NULL;
}
}
// Read the FST
FstReadOptions ropts("<unspecified>", &hdr);
Fst<StdArc> *fst = Fst<StdArc>::Read(ki.Stream(), ropts);
if (!fst) {
if (throw_on_err) {
KALDI_ERR << "Could not read fst from "
<< kaldi::PrintableRxfilename(rxfilename);
} else {
KALDI_WARN << "Could not read fst from "
<< kaldi::PrintableRxfilename(rxfilename)
<< ". A NULL pointer is returned.";
return NULL;
}
}
return fst;
}
VectorFst<StdArc> *CastOrConvertToVectorFst(Fst<StdArc> *fst) {
// This version currently supports ConstFst<StdArc> or VectorFst<StdArc>
std::string real_type = fst->Type();
KALDI_ASSERT(real_type == "vector" || real_type == "const");
if (real_type == "vector") {
return dynamic_cast<VectorFst<StdArc> *>(fst);
} else {
// As the 'fst' can't cast to VectorFst, we create a new
// VectorFst<StdArc> initialized by 'fst', and delete 'fst'.
VectorFst<StdArc> *new_fst = new VectorFst<StdArc>(*fst);
delete fst;
return new_fst;
}
}
void ReadFstKaldi(std::string rxfilename, fst::StdVectorFst *ofst) {
fst::StdVectorFst *fst = ReadFstKaldi(rxfilename);
*ofst = *fst;
delete fst;
}
void WriteFstKaldi(const VectorFst<StdArc> &fst, std::string wxfilename) {
if (wxfilename == "") wxfilename = "-"; // interpret "" as stdout,
// for compatibility with OpenFst conventions.
bool write_binary = true, write_header = false;
kaldi::Output ko(wxfilename, write_binary, write_header);
FstWriteOptions wopts(kaldi::PrintableWxfilename(wxfilename));
fst.Write(ko.Stream(), wopts);
}
fst::VectorFst<fst::StdArc> *ReadAndPrepareLmFst(std::string rxfilename) {
// ReadFstKaldi() will die with exception on failure.
fst::VectorFst<fst::StdArc> *ans = fst::ReadFstKaldi(rxfilename);
if (ans->Properties(fst::kAcceptor, true) == 0) {
// If it's not already an acceptor, project on the output, i.e. copy olabels
// to ilabels. Generally the G.fst's on disk will have the disambiguation
// symbol #0 on the input symbols of the backoff arc, and projection will
// replace them with epsilons which is what is on the output symbols of
// those arcs.
fst::Project(ans, fst::PROJECT_OUTPUT);
}
if (ans->Properties(fst::kILabelSorted, true) == 0) {
// Make sure LM is sorted on ilabel.
fst::ILabelCompare<fst::StdArc> ilabel_comp;
fst::ArcSort(ans, ilabel_comp);
}
return ans;
}
} // end namespace fst
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
此差异已折叠。
add_library(kaldi-lat
determinize-lattice-pruned.cc
lattice-functions.cc
)
target_link_libraries(kaldi-lat PUBLIC kaldi-util)
\ No newline at end of file
...@@ -23,7 +23,7 @@ ...@@ -23,7 +23,7 @@
#include "fstext/fstext-lib.h" #include "fstext/fstext-lib.h"
#include "base/kaldi-common.h" #include "base/kaldi-common.h"
#include "util/common-utils.h" // #include "util/common-utils.h"
namespace kaldi { namespace kaldi {
...@@ -142,13 +142,13 @@ class LatticeHolder { ...@@ -142,13 +142,13 @@ class LatticeHolder {
T *t_; T *t_;
}; };
typedef TableWriter<LatticeHolder> LatticeWriter; // typedef TableWriter<LatticeHolder> LatticeWriter;
typedef SequentialTableReader<LatticeHolder> SequentialLatticeReader; // typedef SequentialTableReader<LatticeHolder> SequentialLatticeReader;
typedef RandomAccessTableReader<LatticeHolder> RandomAccessLatticeReader; // typedef RandomAccessTableReader<LatticeHolder> RandomAccessLatticeReader;
//
typedef TableWriter<CompactLatticeHolder> CompactLatticeWriter; // typedef TableWriter<CompactLatticeHolder> CompactLatticeWriter;
typedef SequentialTableReader<CompactLatticeHolder> SequentialCompactLatticeReader; // typedef SequentialTableReader<CompactLatticeHolder> SequentialCompactLatticeReader;
typedef RandomAccessTableReader<CompactLatticeHolder> RandomAccessCompactLatticeReader; // typedef RandomAccessTableReader<CompactLatticeHolder> RandomAccessCompactLatticeReader;
} // namespace kaldi } // namespace kaldi
......
此差异已折叠。
此差异已折叠。
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册