提交 b3339631 编写于 作者: Y Yibing Liu

Reimpl in class to sepearte init and decoding

上级 1841cea1
......@@ -13,153 +13,146 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "post_decode_faster.h"
#include "base/kaldi-common.h"
#include "base/timer.h"
#include "decoder/decodable-matrix.h"
#include "decoder/faster-decoder.h"
#include "fstext/fstext-lib.h"
#include "hmm/transition-model.h"
#include "lat/kaldi-lattice.h" // for {Compact}LatticeArc
#include "tree/context-dep.h"
#include "util/common-utils.h"
std::vector<std::string> decode(std::string word_syms_filename,
std::string fst_in_filename,
std::string logprior_rxfilename,
std::string posterior_rspecifier,
std::string words_wspecifier,
std::string alignment_wspecifier) {
using namespace kaldi;
typedef kaldi::int32 int32;
using fst::SymbolTable;
using fst::VectorFst;
using fst::StdArc;
Decoder::Decoder(std::string word_syms_filename,
std::string fst_in_filename,
std::string logprior_rxfilename) {
const char *usage =
"Decode, reading log-likelihoods (of transition-ids or whatever symbol "
"is on the graph) as matrices.";
std::string words_wspecifier = "ark,t:out.ark";
std::string alignment_wspecifier = "";
ParseOptions po(usage);
binary = true;
acoustic_scale = 1.5;
allow_partial = true;
FasterDecoderOptions decoder_opts;
decoder_opts.Register(&po, true); // true == include obscure settings.
po.Register("binary", &binary, "Write output in binary mode");
po.Register("allow-partial",
&allow_partial,
"Produce output even when final state was not reached");
po.Register("acoustic-scale",
&acoustic_scale,
"Scaling factor for acoustic likelihoods");
words_writer = new Int32VectorWriter(words_wspecifier);
alignment_writer = new Int32VectorWriter(alignment_wspecifier);
word_syms = NULL;
if (word_syms_filename != "") {
word_syms = fst::SymbolTable::ReadText(word_syms_filename);
if (!word_syms)
KALDI_ERR << "Could not read symbol table from file "
<< word_syms_filename;
}
std::ifstream is_logprior(logprior_rxfilename);
logprior.Read(is_logprior, false);
// It's important that we initialize decode_fst after loglikes_reader, as it
// can prevent crashes on systems installed without enough virtual memory.
// It has to do with what happens on UNIX systems if you call fork() on a
// large process: the page-table entries are duplicated, which requires a
// lot of virtual memory.
decode_fst = fst::ReadFstKaldi(fst_in_filename);
decoder = new FasterDecoder(*decode_fst, decoder_opts);
}
Decoder::~Decoder() {
if (!word_syms) delete word_syms;
delete decode_fst;
delete decoder;
delete words_writer;
delete alignment_writer;
}
std::vector<std::string> Decoder::decode(std::string posterior_rspecifier) {
SequentialBaseFloatMatrixReader posterior_reader(posterior_rspecifier);
std::vector<std::string> decoding_results;
try {
using namespace kaldi;
typedef kaldi::int32 int32;
using fst::SymbolTable;
using fst::VectorFst;
using fst::StdArc;
const char *usage =
"Decode, reading log-likelihoods (of transition-ids or whatever symbol "
"is on the graph) as matrices.";
ParseOptions po(usage);
bool binary = true;
BaseFloat acoustic_scale = 1.5;
bool allow_partial = true;
FasterDecoderOptions decoder_opts;
decoder_opts.Register(&po, true); // true == include obscure settings.
po.Register("binary", &binary, "Write output in binary mode");
po.Register("allow-partial",
&allow_partial,
"Produce output even when final state was not reached");
po.Register("acoustic-scale",
&acoustic_scale,
"Scaling factor for acoustic likelihoods");
Int32VectorWriter words_writer(words_wspecifier);
Int32VectorWriter alignment_writer(alignment_wspecifier);
fst::SymbolTable *word_syms = NULL;
if (word_syms_filename != "") {
word_syms = fst::SymbolTable::ReadText(word_syms_filename);
if (!word_syms)
KALDI_ERR << "Could not read symbol table from file "
<< word_syms_filename;
BaseFloat tot_like = 0.0;
kaldi::int64 frame_count = 0;
int num_success = 0, num_fail = 0;
Timer timer;
for (; !posterior_reader.Done(); posterior_reader.Next()) {
std::string key = posterior_reader.Key();
Matrix<BaseFloat> loglikes(posterior_reader.Value());
KALDI_LOG << key << " " << loglikes.NumRows() << " x "
<< loglikes.NumCols();
if (loglikes.NumRows() == 0) {
KALDI_WARN << "Zero-length utterance: " << key;
num_fail++;
continue;
}
SequentialBaseFloatMatrixReader posterior_reader(posterior_rspecifier);
std::ifstream is_logprior(logprior_rxfilename);
Vector<BaseFloat> logprior;
logprior.Read(is_logprior, false);
// It's important that we initialize decode_fst after loglikes_reader, as it
// can prevent crashes on systems installed without enough virtual memory.
// It has to do with what happens on UNIX systems if you call fork() on a
// large process: the page-table entries are duplicated, which requires a
// lot of virtual memory.
VectorFst<StdArc> *decode_fst = fst::ReadFstKaldi(fst_in_filename);
BaseFloat tot_like = 0.0;
kaldi::int64 frame_count = 0;
int num_success = 0, num_fail = 0;
FasterDecoder decoder(*decode_fst, decoder_opts);
Timer timer;
for (; !posterior_reader.Done(); posterior_reader.Next()) {
std::string key = posterior_reader.Key();
Matrix<BaseFloat> loglikes(posterior_reader.Value());
KALDI_LOG << key << " " << loglikes.NumRows() << " x "
<< loglikes.NumCols();
if (loglikes.NumRows() == 0) {
KALDI_WARN << "Zero-length utterance: " << key;
num_fail++;
continue;
}
KALDI_ASSERT(loglikes.NumCols() == logprior.Dim());
loglikes.ApplyLog();
loglikes.AddVecToRows(-1.0, logprior);
DecodableMatrixScaled decodable(loglikes, acoustic_scale);
decoder.Decode(&decodable);
VectorFst<LatticeArc> decoded; // linear FST.
if ((allow_partial || decoder.ReachedFinal()) &&
decoder.GetBestPath(&decoded)) {
num_success++;
if (!decoder.ReachedFinal())
KALDI_WARN << "Decoder did not reach end-state, outputting partial "
"traceback.";
std::vector<int32> alignment;
std::vector<int32> words;
LatticeWeight weight;
frame_count += loglikes.NumRows();
GetLinearSymbolSequence(decoded, &alignment, &words, &weight);
words_writer.Write(key, words);
if (alignment_writer.IsOpen()) alignment_writer.Write(key, alignment);
if (word_syms != NULL) {
std::string res;
for (size_t i = 0; i < words.size(); i++) {
std::string s = word_syms->Find(words[i]);
res += s;
if (s == "")
KALDI_ERR << "Word-id " << words[i] << " not in symbol table.";
std::cerr << s << ' ';
}
decoding_results.push_back(res);
KALDI_ASSERT(loglikes.NumCols() == logprior.Dim());
loglikes.ApplyLog();
loglikes.AddVecToRows(-1.0, logprior);
DecodableMatrixScaled decodable(loglikes, acoustic_scale);
decoder->Decode(&decodable);
VectorFst<LatticeArc> decoded; // linear FST.
if ((allow_partial || decoder->ReachedFinal()) &&
decoder->GetBestPath(&decoded)) {
num_success++;
if (!decoder->ReachedFinal())
KALDI_WARN << "Decoder did not reach end-state, outputting partial "
"traceback.";
std::vector<int32> alignment;
std::vector<int32> words;
LatticeWeight weight;
frame_count += loglikes.NumRows();
GetLinearSymbolSequence(decoded, &alignment, &words, &weight);
words_writer->Write(key, words);
if (alignment_writer->IsOpen()) alignment_writer->Write(key, alignment);
if (word_syms != NULL) {
std::string res;
for (size_t i = 0; i < words.size(); i++) {
std::string s = word_syms->Find(words[i]);
res += s;
if (s == "")
KALDI_ERR << "Word-id " << words[i] << " not in symbol table.";
std::cerr << s << ' ';
}
BaseFloat like = -weight.Value1() - weight.Value2();
tot_like += like;
KALDI_LOG << "Log-like per frame for utterance " << key << " is "
<< (like / loglikes.NumRows()) << " over "
<< loglikes.NumRows() << " frames.";
} else {
num_fail++;
KALDI_WARN << "Did not successfully decode utterance " << key
<< ", len = " << loglikes.NumRows();
decoding_results.push_back(res);
}
BaseFloat like = -weight.Value1() - weight.Value2();
tot_like += like;
KALDI_LOG << "Log-like per frame for utterance " << key << " is "
<< (like / loglikes.NumRows()) << " over " << loglikes.NumRows()
<< " frames.";
} else {
num_fail++;
KALDI_WARN << "Did not successfully decode utterance " << key
<< ", len = " << loglikes.NumRows();
}
double elapsed = timer.Elapsed();
KALDI_LOG << "Time taken [excluding initialization] " << elapsed
<< "s: real-time factor assuming 100 frames/sec is "
<< (elapsed * 100.0 / frame_count);
KALDI_LOG << "Done " << num_success << " utterances, failed for "
<< num_fail;
KALDI_LOG << "Overall log-likelihood per frame is "
<< (tot_like / frame_count) << " over " << frame_count
<< " frames.";
delete word_syms;
delete decode_fst;
} catch (const std::exception &e) {
std::cerr << e.what();
}
double elapsed = timer.Elapsed();
KALDI_LOG << "Time taken [excluding initialization] " << elapsed
<< "s: real-time factor assuming 100 frames/sec is "
<< (elapsed * 100.0 / frame_count);
KALDI_LOG << "Done " << num_success << " utterances, failed for " << num_fail;
KALDI_LOG << "Overall log-likelihood per frame is "
<< (tot_like / frame_count) << " over " << frame_count
<< " frames.";
return decoding_results;
}
......@@ -14,10 +14,36 @@ limitations under the License. */
#include <string>
#include <vector>
#include "base/kaldi-common.h"
#include "base/timer.h"
#include "decoder/decodable-matrix.h"
#include "decoder/faster-decoder.h"
#include "fstext/fstext-lib.h"
#include "hmm/transition-model.h"
#include "lat/kaldi-lattice.h" // for {Compact}LatticeArc
#include "tree/context-dep.h"
#include "util/common-utils.h"
std::vector<std::string> decode(std::string word_syms_filename,
std::string fst_in_filename,
std::string logprior_rxfilename,
std::string posterior_respecifier,
std::string words_wspecifier,
std::string alignment_wspecifier = "");
class Decoder {
public:
Decoder(std::string word_syms_filename,
std::string fst_in_filename,
std::string logprior_rxfilename);
~Decoder();
std::vector<std::string> decode(std::string posterior_rspecifier);
private:
fst::SymbolTable *word_syms;
fst::VectorFst<fst::StdArc> *decode_fst;
kaldi::FasterDecoder *decoder;
kaldi::Vector<kaldi::BaseFloat> logprior;
kaldi::Int32VectorWriter *words_writer;
kaldi::Int32VectorWriter *alignment_writer;
bool binary;
kaldi::BaseFloat acoustic_scale;
bool allow_partial;
};
......@@ -20,10 +20,12 @@ limitations under the License. */
namespace py = pybind11;
PYBIND11_MODULE(post_decode_faster, m) {
m.doc() = "Decode function for Deep ASR model";
m.def("decode",
&decode,
"Decode one input probability matrix "
"and return the transcription");
m.doc() = "Decoder for Deep ASR model";
py::class_<Decoder>(m, "Decoder")
.def(py::init<std::string, std::string, std::string>())
.def("decode",
&Decoder::decode,
"Decode one input probability matrix "
"and return the transcription");
}
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