EngineX-Iluvatar/enginex_bi_series-sherpa-onnx
Archived
16
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

Added provider option to sherpa-onnx and decode-file-c-api (#162)

Browse Source
This commit is contained in:
Jingzhao Ou
2023-06-02 13:57:48 -07:00
committed by GitHub
parent 5e2dc5ceea
commit 0ed501b8f1
9 changed files with 265 additions and 171 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
sherpa-onnx/c-api/c-api.cc
View File

@@ -41,6 +41,7 @@ SherpaOnnxOnlineRecognizer *CreateOnlineRecognizer(
recognizer_config.model_config.joiner_filename = config->model_config.joiner;
recognizer_config.model_config.tokens = config->model_config.tokens;
recognizer_config.model_config.num_threads = config->model_config.num_threads;
recognizer_config.model_config.provider = config->model_config.provider;
recognizer_config.model_config.debug = config->model_config.debug;
recognizer_config.decoding_method = config->decoding_method;

1
sherpa-onnx/c-api/c-api.h
View File

@@ -52,6 +52,7 @@ SHERPA_ONNX_API typedef struct SherpaOnnxOnlineTransducerModelConfig {
const char *joiner;
const char *tokens;
int32_t num_threads;
const char *provider;
int32_t debug; // true to print debug information of the model
} SherpaOnnxOnlineTransducerModelConfig;

3
sherpa-onnx/csrc/online-transducer-model-config.cc
View File

@@ -17,6 +17,8 @@ void OnlineTransducerModelConfig::Register(ParseOptions *po) {
po->Register("tokens", &tokens, "Path to tokens.txt");
po->Register("num_threads", &num_threads,
"Number of threads to run the neural network");
po->Register("provider", &provider,
"Specify a provider to use: cpu, cuda, coreml");
po->Register("debug", &debug,
"true to print model information while loading it.");
@@ -60,6 +62,7 @@ std::string OnlineTransducerModelConfig::ToString() const {
os << "joiner_filename=\"" << joiner_filename << "\", ";
os << "tokens=\"" << tokens << "\", ";
os << "num_threads=" << num_threads << ", ";
os << "provider=\"" << provider << "\", ";
os << "debug=" << (debug ? "True" : "False") << ")";
return os.str();

8
sherpa-onnx/csrc/sherpa-onnx-offline.cc
View File

@@ -69,17 +69,17 @@ for a list of pre-trained models to download.
fprintf(stderr, "Creating recognizer ...\n");
sherpa_onnx::OfflineRecognizer recognizer(config);
auto begin = std::chrono::steady_clock::now();
const auto begin = std::chrono::steady_clock::now();
fprintf(stderr, "Started\n");
std::vector<std::unique_ptr<sherpa_onnx::OfflineStream>> ss;
std::vector<sherpa_onnx::OfflineStream *> ss_pointers;
float duration = 0;
for (int32_t i = 1; i <= po.NumArgs(); ++i) {
std::string wav_filename = po.GetArg(i);
const std::string wav_filename = po.GetArg(i);
int32_t sampling_rate = -1;
bool is_ok = false;
std::vector<float> samples =
const std::vector<float> samples =
sherpa_onnx::ReadWave(wav_filename, &sampling_rate, &is_ok);
if (!is_ok) {
fprintf(stderr, "Failed to read %s\n", wav_filename.c_str());
@@ -96,7 +96,7 @@ for a list of pre-trained models to download.
recognizer.DecodeStreams(ss_pointers.data(), ss_pointers.size());
auto end = std::chrono::steady_clock::now();
const auto end = std::chrono::steady_clock::now();
fprintf(stderr, "Done!\n\n");
for (int32_t i = 1; i <= po.NumArgs(); ++i) {

161
sherpa-onnx/csrc/sherpa-onnx.cc
View File

@@ -11,22 +11,28 @@
#include "sherpa-onnx/csrc/online-recognizer.h"
#include "sherpa-onnx/csrc/online-stream.h"
#include "sherpa-onnx/csrc/symbol-table.h"
#include "sherpa-onnx/csrc/parse-options.h"
#include "sherpa-onnx/csrc/wave-reader.h"
// TODO(fangjun): Use ParseOptions as we are getting more args
int main(int32_t argc, char *argv[]) {
if (argc < 6 || argc > 9) {
const char *usage = R"usage(
const char *kUsageMessage = R"usage(
Usage:
./bin/sherpa-onnx \
/path/to/tokens.txt \
/path/to/encoder.onnx \
/path/to/decoder.onnx \
/path/to/joiner.onnx \
/path/to/foo.wav [num_threads [decoding_method [/path/to/rnn_lm.onnx]]]
--tokens=/path/to/tokens.txt \
--encoder=/path/to/encoder.onnx \
--decoder=/path/to/decoder.onnx \
--joiner=/path/to/joiner.onnx \
--provider=cpu \
--num-threads=2 \
--decoding-method=greedy_search \
/path/to/foo.wav [bar.wav foobar.wav ...]
Note: It supports decoding multiple files in batches
Default value for num_threads is 2.
Valid values for decoding_method: greedy_search (default), modified_beam_search.
Valid values for provider: cpu (default), cuda, coreml.
foo.wav should be of single channel, 16-bit PCM encoded wave file; its
sampling rate can be arbitrary and does not need to be 16kHz.
@@ -34,33 +40,17 @@ Please refer to
https://k2-fsa.github.io/sherpa/onnx/pretrained_models/index.html
for a list of pre-trained models to download.
)usage";
fprintf(stderr, "%s\n", usage);
return 0;
}
sherpa_onnx::ParseOptions po(kUsageMessage);
sherpa_onnx::OnlineRecognizerConfig config;
config.model_config.tokens = argv[1];
config.Register(&po);
config.model_config.debug = false;
config.model_config.encoder_filename = argv[2];
config.model_config.decoder_filename = argv[3];
config.model_config.joiner_filename = argv[4];
std::string wav_filename = argv[5];
config.model_config.num_threads = 2;
if (argc == 7 && atoi(argv[6]) > 0) {
config.model_config.num_threads = atoi(argv[6]);
po.Read(argc, argv);
if (po.NumArgs() < 1) {
po.PrintUsage();
exit(EXIT_FAILURE);
}
if (argc == 8) {
config.decoding_method = argv[7];
}
if (argc == 9) {
config.lm_config.model = argv[8];
}
config.max_active_paths = 4;
fprintf(stderr, "%s\n", config.ToString().c_str());
@@ -71,63 +61,66 @@ for a list of pre-trained models to download.
sherpa_onnx::OnlineRecognizer recognizer(config);
int32_t sampling_rate = -1;
float duration = 0;
for (int32_t i = 1; i <= po.NumArgs(); ++i) {
const std::string wav_filename = po.GetArg(i);
int32_t sampling_rate = -1;
bool is_ok = false;
std::vector<float> samples =
sherpa_onnx::ReadWave(wav_filename, &sampling_rate, &is_ok);
bool is_ok = false;
const std::vector<float> samples =
sherpa_onnx::ReadWave(wav_filename, &sampling_rate, &is_ok);
if (!is_ok) {
fprintf(stderr, "Failed to read %s\n", wav_filename.c_str());
return -1;
if (!is_ok) {
fprintf(stderr, "Failed to read %s\n", wav_filename.c_str());
return -1;
}
fprintf(stderr, "sampling rate of input file: %d\n", sampling_rate);
const float duration = samples.size() / static_cast<float>(sampling_rate);
fprintf(stderr, "wav filename: %s\n", wav_filename.c_str());
fprintf(stderr, "wav duration (s): %.3f\n", duration);
fprintf(stderr, "Started\n");
const auto begin = std::chrono::steady_clock::now();
auto s = recognizer.CreateStream();
s->AcceptWaveform(sampling_rate, samples.data(), samples.size());
std::vector<float> tail_paddings(static_cast<int>(0.3 * sampling_rate));
// Note: We can call AcceptWaveform() multiple times.
s->AcceptWaveform(
sampling_rate, tail_paddings.data(), tail_paddings.size());
// Call InputFinished() to indicate that no audio samples are available
s->InputFinished();
while (recognizer.IsReady(s.get())) {
recognizer.DecodeStream(s.get());
}
const std::string text = recognizer.GetResult(s.get()).AsJsonString();
const auto end = std::chrono::steady_clock::now();
const float elapsed_seconds =
std::chrono::duration_cast<std::chrono::milliseconds>(end - begin)
.count() / 1000.;
fprintf(stderr, "Done!\n");
fprintf(stderr,
"Recognition result for %s:\n%s\n",
wav_filename.c_str(), text.c_str());
fprintf(stderr, "num threads: %d\n", config.model_config.num_threads);
fprintf(stderr, "decoding method: %s\n", config.decoding_method.c_str());
if (config.decoding_method == "modified_beam_search") {
fprintf(stderr, "max active paths: %d\n", config.max_active_paths);
}
fprintf(stderr, "Elapsed seconds: %.3f s\n", elapsed_seconds);
const float rtf = elapsed_seconds / duration;
fprintf(stderr, "Real time factor (RTF): %.3f / %.3f = %.3f\n",
elapsed_seconds, duration, rtf);
}
fprintf(stderr, "sampling rate of input file: %d\n", sampling_rate);
float duration = samples.size() / static_cast<float>(sampling_rate);
fprintf(stderr, "wav filename: %s\n", wav_filename.c_str());
fprintf(stderr, "wav duration (s): %.3f\n", duration);
auto begin = std::chrono::steady_clock::now();
fprintf(stderr, "Started\n");
auto s = recognizer.CreateStream();
s->AcceptWaveform(sampling_rate, samples.data(), samples.size());
std::vector<float> tail_paddings(static_cast<int>(0.3 * sampling_rate));
// Note: We can call AcceptWaveform() multiple times.
s->AcceptWaveform(sampling_rate, tail_paddings.data(), tail_paddings.size());
// Call InputFinished() to indicate that no audio samples are available
s->InputFinished();
while (recognizer.IsReady(s.get())) {
recognizer.DecodeStream(s.get());
}
std::string text = recognizer.GetResult(s.get()).AsJsonString();
fprintf(stderr, "Done!\n");
fprintf(stderr, "Recognition result for %s:\n%s\n", wav_filename.c_str(),
text.c_str());
auto end = std::chrono::steady_clock::now();
float elapsed_seconds =
std::chrono::duration_cast<std::chrono::milliseconds>(end - begin)
.count() /
1000.;
fprintf(stderr, "num threads: %d\n", config.model_config.num_threads);
fprintf(stderr, "decoding method: %s\n", config.decoding_method.c_str());
if (config.decoding_method == "modified_beam_search") {
fprintf(stderr, "max active paths: %d\n", config.max_active_paths);
}
fprintf(stderr, "Elapsed seconds: %.3f s\n", elapsed_seconds);
float rtf = elapsed_seconds / duration;
fprintf(stderr, "Real time factor (RTF): %.3f / %.3f = %.3f\n",
elapsed_seconds, duration, rtf);
return 0;
}