add paraformer-torch

雾聪

2024-03-13 7675a2a0baa30357da00263186964c0d0d814581

add paraformer-torch

 runtime/onnxruntime/include/funasrruntime.h

@@ -96,7 +96,7 @@
_FUNASRAPI void                    CTTransformerUninit(FUNASR_HANDLE handle);

//OfflineStream
_FUNASRAPI FUNASR_HANDLE      FunOfflineInit(std::map<std::string, std::string>& model_path, int thread_num);
_FUNASRAPI FUNASR_HANDLE      FunOfflineInit(std::map<std::string, std::string>& model_path, int thread_num, bool use_gpu=false);
_FUNASRAPI void             FunOfflineReset(FUNASR_HANDLE handle, FUNASR_DEC_HANDLE dec_handle=nullptr);
// buffer
_FUNASRAPI FUNASR_RESULT    FunOfflineInferBuffer(FUNASR_HANDLE handle, const char* sz_buf, int n_len, 

 runtime/onnxruntime/include/offline-stream.h

@@ -14,7 +14,7 @@
namespace funasr {
class OfflineStream {
  public:
    OfflineStream(std::map<std::string, std::string>& model_path, int thread_num);
    OfflineStream(std::map<std::string, std::string>& model_path, int thread_num, bool use_gpu=false);
    ~OfflineStream(){};

    std::unique_ptr<VadModel> vad_handle= nullptr;
@@ -33,6 +33,6 @@
    bool use_itn=false;
};

OfflineStream *CreateOfflineStream(std::map<std::string, std::string>& model_path, int thread_num=1);
OfflineStream *CreateOfflineStream(std::map<std::string, std::string>& model_path, int thread_num=1, bool use_gpu=false);
} // namespace funasr
#endif

 runtime/onnxruntime/src/CMakeLists.txt

@@ -25,7 +25,11 @@
    include_directories(${FFMPEG_DIR}/include)
endif()

if(GPU)
    set(TORCH_DEPS torch torch_cuda torch_cpu c10 c10_cuda torch_blade ral_base_context)
endif()

#message("CXX_FLAGS "${CMAKE_CXX_FLAGS})
include_directories(${CMAKE_SOURCE_DIR}/include)
include_directories(${CMAKE_SOURCE_DIR}/third_party)
target_link_libraries(funasr PUBLIC onnxruntime ${EXTRA_LIBS})
target_link_libraries(funasr PUBLIC onnxruntime ${EXTRA_LIBS} ${TORCH_DEPS})

 runtime/onnxruntime/src/funasrruntime.cpp

@@ -33,9 +33,9 @@
        return mm;
    }

    _FUNASRAPI FUNASR_HANDLE  FunOfflineInit(std::map<std::string, std::string>& model_path, int thread_num)
    _FUNASRAPI FUNASR_HANDLE  FunOfflineInit(std::map<std::string, std::string>& model_path, int thread_num, bool use_gpu)
    {
        funasr::OfflineStream* mm = funasr::CreateOfflineStream(model_path, thread_num);
        funasr::OfflineStream* mm = funasr::CreateOfflineStream(model_path, thread_num, use_gpu);
        return mm;
    }


 runtime/onnxruntime/src/offline-stream.cpp

@@ -1,7 +1,7 @@
#include "precomp.h"

namespace funasr {
OfflineStream::OfflineStream(std::map<std::string, std::string>& model_path, int thread_num)
OfflineStream::OfflineStream(std::map<std::string, std::string>& model_path, int thread_num, bool use_gpu)
{
    // VAD model
    if(model_path.find(VAD_DIR) != model_path.end()){
@@ -35,7 +35,12 @@
        string hw_compile_model_path;
        string seg_dict_path;
    
        asr_handle = make_unique<Paraformer>();
        if(use_gpu){
            asr_handle = make_unique<ParaformerTorch>();
        }else{
            asr_handle = make_unique<Paraformer>();
        }

        bool enable_hotword = false;
        hw_compile_model_path = PathAppend(model_path.at(MODEL_DIR), MODEL_EB_NAME);
        seg_dict_path = PathAppend(model_path.at(MODEL_DIR), MODEL_SEG_DICT);
@@ -115,10 +120,10 @@
#endif
}

OfflineStream *CreateOfflineStream(std::map<std::string, std::string>& model_path, int thread_num)
OfflineStream *CreateOfflineStream(std::map<std::string, std::string>& model_path, int thread_num, bool use_gpu)
{
    OfflineStream *mm;
    mm = new OfflineStream(model_path, thread_num);
    mm = new OfflineStream(model_path, thread_num, use_gpu);
    return mm;
}


 runtime/onnxruntime/src/paraformer-torch.cpp

New file
@@ -0,0 +1,351 @@
/**
 * Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
 * MIT License  (https://opensource.org/licenses/MIT)
*/

#include "precomp.h"
#include "paraformer-torch.h"
#include "encode_converter.h"
#include <cstddef>

using namespace std;
namespace funasr {

ParaformerTorch::ParaformerTorch()
:use_hotword(false){
}

// offline
void ParaformerTorch::InitAsr(const std::string &am_model, const std::string &am_cmvn, const std::string &am_config, int thread_num){
    LoadConfigFromYaml(am_config.c_str());
    // knf options
    fbank_opts_.frame_opts.dither = 0;
    fbank_opts_.mel_opts.num_bins = n_mels;
    fbank_opts_.frame_opts.samp_freq = asr_sample_rate;
    fbank_opts_.frame_opts.window_type = window_type;
    fbank_opts_.frame_opts.frame_shift_ms = frame_shift;
    fbank_opts_.frame_opts.frame_length_ms = frame_length;
    fbank_opts_.energy_floor = 0;
    fbank_opts_.mel_opts.debug_mel = false;

    vocab = new Vocab(am_config.c_str());
    phone_set_ = new PhoneSet(am_config.c_str());
    LoadCmvn(am_cmvn.c_str());

    torch::DeviceType device = at::kCPU;
    #ifdef USE_GPU
    if (!torch::cuda::is_available()) {
        LOG(ERROR) << "CUDA is not available! Please check your GPU settings";
        exit(-1);
    } else {
        LOG(INFO) << "CUDA available! Running on GPU";
        device = at::kCUDA;
    }
    #endif
    #ifdef USE_IPEX
    torch::jit::setTensorExprFuserEnabled(false);
    #endif
    torch::jit::script::Module model = torch::jit::load(am_model, device);
    model_ = std::make_shared<TorchModule>(std::move(model));    
}

void ParaformerTorch::InitLm(const std::string &lm_file, 
                        const std::string &lm_cfg_file, 
                        const std::string &lex_file) {
    try {
        lm_ = std::shared_ptr<fst::Fst<fst::StdArc>>(
            fst::Fst<fst::StdArc>::Read(lm_file));
        if (lm_){
            lm_vocab = new Vocab(lm_cfg_file.c_str(), lex_file.c_str());
            LOG(INFO) << "Successfully load lm file " << lm_file;
        }else{
            LOG(ERROR) << "Failed to load lm file " << lm_file;
        }
    } catch (std::exception const &e) {
        LOG(ERROR) << "Error when load lm file: " << e.what();
        exit(0);
    }
}

void ParaformerTorch::LoadConfigFromYaml(const char* filename){

    YAML::Node config;
    try{
        config = YAML::LoadFile(filename);
    }catch(exception const &e){
        LOG(ERROR) << "Error loading file, yaml file error or not exist.";
        exit(-1);
    }

    try{
        YAML::Node frontend_conf = config["frontend_conf"];
        this->asr_sample_rate = frontend_conf["fs"].as<int>();

        YAML::Node lang_conf = config["lang"];
        if (lang_conf.IsDefined()){
            language = lang_conf.as<string>();
        }
    }catch(exception const &e){
        LOG(ERROR) << "Error when load argument from vad config YAML.";
        exit(-1);
    }
}

void ParaformerTorch::InitHwCompiler(const std::string &hw_model, int thread_num) {
    // TODO
    use_hotword = true;
}

void ParaformerTorch::InitSegDict(const std::string &seg_dict_model) {
    seg_dict = new SegDict(seg_dict_model.c_str());
}

ParaformerTorch::~ParaformerTorch()
{
    if(vocab){
        delete vocab;
    }
    if(lm_vocab){
        delete lm_vocab;
    }
    if(seg_dict){
        delete seg_dict;
    }
    if(phone_set_){
        delete phone_set_;
    }
}

void ParaformerTorch::StartUtterance()
{
}

void ParaformerTorch::EndUtterance()
{
}

void ParaformerTorch::Reset()
{
}

void ParaformerTorch::FbankKaldi(float sample_rate, const float* waves, int len, std::vector<std::vector<float>> &asr_feats) {
    knf::OnlineFbank fbank_(fbank_opts_);
    std::vector<float> buf(len);
    for (int32_t i = 0; i != len; ++i) {
        buf[i] = waves[i] * 32768;
    }
    fbank_.AcceptWaveform(sample_rate, buf.data(), buf.size());

    int32_t frames = fbank_.NumFramesReady();
    for (int32_t i = 0; i != frames; ++i) {
        const float *frame = fbank_.GetFrame(i);
        std::vector<float> frame_vector(frame, frame + fbank_opts_.mel_opts.num_bins);
        asr_feats.emplace_back(frame_vector);
    }
}

void ParaformerTorch::LoadCmvn(const char *filename)
{
    ifstream cmvn_stream(filename);
    if (!cmvn_stream.is_open()) {
        LOG(ERROR) << "Failed to open file: " << filename;
        exit(-1);
    }
    string line;

    while (getline(cmvn_stream, line)) {
        istringstream iss(line);
        vector<string> line_item{istream_iterator<string>{iss}, istream_iterator<string>{}};
        if (line_item[0] == "<AddShift>") {
            getline(cmvn_stream, line);
            istringstream means_lines_stream(line);
            vector<string> means_lines{istream_iterator<string>{means_lines_stream}, istream_iterator<string>{}};
            if (means_lines[0] == "<LearnRateCoef>") {
                for (int j = 3; j < means_lines.size() - 1; j++) {
                    means_list_.push_back(stof(means_lines[j]));
                }
                continue;
            }
        }
        else if (line_item[0] == "<Rescale>") {
            getline(cmvn_stream, line);
            istringstream vars_lines_stream(line);
            vector<string> vars_lines{istream_iterator<string>{vars_lines_stream}, istream_iterator<string>{}};
            if (vars_lines[0] == "<LearnRateCoef>") {
                for (int j = 3; j < vars_lines.size() - 1; j++) {
                    vars_list_.push_back(stof(vars_lines[j])*scale);
                }
                continue;
            }
        }
    }
}

string ParaformerTorch::GreedySearch(float * in, int n_len,  int64_t token_nums, bool is_stamp, std::vector<float> us_alphas, std::vector<float> us_cif_peak)
{
    vector<int> hyps;
    int Tmax = n_len;
    for (int i = 0; i < Tmax; i++) {
        int max_idx;
        float max_val;
        FindMax(in + i * token_nums, token_nums, max_val, max_idx);
        hyps.push_back(max_idx);
    }
    if(!is_stamp){
        return vocab->Vector2StringV2(hyps, language);
    }else{
        std::vector<string> char_list;
        std::vector<std::vector<float>> timestamp_list;
        std::string res_str;
        vocab->Vector2String(hyps, char_list);
        std::vector<string> raw_char(char_list);
        TimestampOnnx(us_alphas, us_cif_peak, char_list, res_str, timestamp_list);

        return PostProcess(raw_char, timestamp_list);
    }
}

string ParaformerTorch::BeamSearch(WfstDecoder* &wfst_decoder, float *in, int len, int64_t token_nums)
{
  return wfst_decoder->Search(in, len, token_nums);
}

string ParaformerTorch::FinalizeDecode(WfstDecoder* &wfst_decoder,
                                  bool is_stamp, std::vector<float> us_alphas, std::vector<float> us_cif_peak)
{
  return wfst_decoder->FinalizeDecode(is_stamp, us_alphas, us_cif_peak);
}

void ParaformerTorch::LfrCmvn(std::vector<std::vector<float>> &asr_feats) {

    std::vector<std::vector<float>> out_feats;
    int T = asr_feats.size();
    int T_lrf = ceil(1.0 * T / lfr_n);

    // Pad frames at start(copy first frame)
    for (int i = 0; i < (lfr_m - 1) / 2; i++) {
        asr_feats.insert(asr_feats.begin(), asr_feats[0]);
    }
    // Merge lfr_m frames as one,lfr_n frames per window
    T = T + (lfr_m - 1) / 2;
    std::vector<float> p;
    for (int i = 0; i < T_lrf; i++) {
        if (lfr_m <= T - i * lfr_n) {
            for (int j = 0; j < lfr_m; j++) {
                p.insert(p.end(), asr_feats[i * lfr_n + j].begin(), asr_feats[i * lfr_n + j].end());
            }
            out_feats.emplace_back(p);
            p.clear();
        } else {
            // Fill to lfr_m frames at last window if less than lfr_m frames  (copy last frame)
            int num_padding = lfr_m - (T - i * lfr_n);
            for (int j = 0; j < (asr_feats.size() - i * lfr_n); j++) {
                p.insert(p.end(), asr_feats[i * lfr_n + j].begin(), asr_feats[i * lfr_n + j].end());
            }
            for (int j = 0; j < num_padding; j++) {
                p.insert(p.end(), asr_feats[asr_feats.size() - 1].begin(), asr_feats[asr_feats.size() - 1].end());
            }
            out_feats.emplace_back(p);
            p.clear();
        }
    }
    // Apply cmvn
    for (auto &out_feat: out_feats) {
        for (int j = 0; j < means_list_.size(); j++) {
            out_feat[j] = (out_feat[j] + means_list_[j]) * vars_list_[j];
        }
    }
    asr_feats = out_feats;
}

string ParaformerTorch::Forward(float* din, int len, bool input_finished, const std::vector<std::vector<float>> &hw_emb, void* decoder_handle)
{
    WfstDecoder* wfst_decoder = (WfstDecoder*)decoder_handle;
    int32_t in_feat_dim = fbank_opts_.mel_opts.num_bins;

    std::vector<std::vector<float>> asr_feats;
    FbankKaldi(asr_sample_rate, din, len, asr_feats);
    if(asr_feats.size() == 0){
      return "";
    }
    LfrCmvn(asr_feats);
    int32_t feat_dim = lfr_m*in_feat_dim;
    int32_t num_frames = asr_feats.size();

    std::vector<float> wav_feats;
    for (const auto &frame_feat: asr_feats) {
        wav_feats.insert(wav_feats.end(), frame_feat.begin(), frame_feat.end());
    }
    std::vector<int32_t> paraformer_length;
    paraformer_length.emplace_back(num_frames);

    torch::NoGradGuard no_grad;
    torch::Tensor feats =
        torch::from_blob(wav_feats.data(),
                {1, num_frames, feat_dim}, torch::kFloat).contiguous();
    torch::Tensor feat_lens = torch::from_blob(paraformer_length.data(),
                        {1}, torch::kInt32);

    // 2. forward
    #ifdef USE_GPU
    feats = feats.to(at::kCUDA);
    feat_lens = feat_lens.to(at::kCUDA);
    #endif
    std::vector<torch::jit::IValue> inputs = {feats, feat_lens};

    string result="";
    try {
        auto outputs = model_->forward(inputs).toTuple()->elements();
        torch::Tensor am_scores;
        torch::Tensor valid_token_lens;
        #ifdef USE_GPU
        am_scores = outputs[0].toTensor().to(at::kCPU);
        valid_token_lens = outputs[1].toTensor().to(at::kCPU);
        #else
        am_scores = outputs[0].toTensor();
        valid_token_lens = outputs[1].toTensor();
        #endif
        
        if (lm_ == nullptr) {
            result = GreedySearch(am_scores[0].data_ptr<float>(), valid_token_lens[0].item<int>(), am_scores.size(2));
        } else {
            result = BeamSearch(wfst_decoder, am_scores[0].data_ptr<float>(), valid_token_lens[0].item<int>(), am_scores.size(2));
            if (input_finished) {
                result = FinalizeDecode(wfst_decoder);
            }
        }        
    }
    catch (std::exception const &e)
    {
        LOG(ERROR)<<e.what();
    }

    return result;
}

std::vector<std::vector<float>> ParaformerTorch::CompileHotwordEmbedding(std::string &hotwords) {
    std::vector<std::vector<float>> result;
    return result;
}

Vocab* ParaformerTorch::GetVocab()
{
    return vocab;
}

Vocab* ParaformerTorch::GetLmVocab()
{
    return lm_vocab;
}

PhoneSet* ParaformerTorch::GetPhoneSet()
{
    return phone_set_;
}

string ParaformerTorch::Rescoring()
{
    LOG(ERROR)<<"Not Imp!!!!!!";
    return "";
}
} // namespace funasr

 runtime/onnxruntime/src/paraformer-torch.h

New file
@@ -0,0 +1,92 @@
/**
 * Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
 * MIT License  (https://opensource.org/licenses/MIT)
*/
#pragma once
#include <torch/serialize.h>
#include <torch/script.h>
#include <torch/torch.h>
#include <torch/csrc/jit/passes/tensorexpr_fuser.h>
#include "precomp.h"
#include "fst/fstlib.h"
#include "fst/symbol-table.h"
#include "bias-lm.h"
#include "phone-set.h"

namespace funasr {

    class ParaformerTorch : public Model {
    /**
     * Author: Speech Lab of DAMO Academy, Alibaba Group
     * Paraformer: Fast and Accurate Parallel Transformer for Non-autoregressive End-to-End Speech Recognition
     * https://arxiv.org/pdf/2206.08317.pdf
    */
    private:
        Vocab* vocab = nullptr;
        Vocab* lm_vocab = nullptr;
        SegDict* seg_dict = nullptr;
        PhoneSet* phone_set_ = nullptr;
        //const float scale = 22.6274169979695;
        const float scale = 1.0;

        void LoadConfigFromYaml(const char* filename);
        void LoadCmvn(const char *filename);
        void LfrCmvn(std::vector<std::vector<float>> &asr_feats);

        using TorchModule = torch::jit::script::Module;
        std::shared_ptr<TorchModule> model_ = nullptr;
        std::vector<torch::Tensor> encoder_outs_;
        bool use_hotword;

    public:
        ParaformerTorch();
        ~ParaformerTorch();
        void InitAsr(const std::string &am_model, const std::string &am_cmvn, const std::string &am_config, int thread_num);
        void InitHwCompiler(const std::string &hw_model, int thread_num);
        void InitSegDict(const std::string &seg_dict_model);
        std::vector<std::vector<float>> CompileHotwordEmbedding(std::string &hotwords);
        void Reset();
        void FbankKaldi(float sample_rate, const float* waves, int len, std::vector<std::vector<float>> &asr_feats);
        string Forward(float* din, int len, bool input_finished=true, const std::vector<std::vector<float>> &hw_emb={{0.0}}, void* wfst_decoder=nullptr);
        string GreedySearch( float* in, int n_len, int64_t token_nums,
                             bool is_stamp=false, std::vector<float> us_alphas={0}, std::vector<float> us_cif_peak={0});

        string Rescoring();
        string GetLang(){return language;};
        int GetAsrSampleRate() { return asr_sample_rate; };
        void StartUtterance();
        void EndUtterance();
        void InitLm(const std::string &lm_file, const std::string &lm_cfg_file, const std::string &lex_file);
        string BeamSearch(WfstDecoder* &wfst_decoder, float* in, int n_len, int64_t token_nums);
        string FinalizeDecode(WfstDecoder* &wfst_decoder,
                          bool is_stamp=false, std::vector<float> us_alphas={0}, std::vector<float> us_cif_peak={0});
        Vocab* GetVocab();
        Vocab* GetLmVocab();
        PhoneSet* GetPhoneSet();
		
        knf::FbankOptions fbank_opts_;
        vector<float> means_list_;
        vector<float> vars_list_;
        int lfr_m = PARA_LFR_M;
        int lfr_n = PARA_LFR_N;

        // paraformer-offline
        std::string language="zh-cn";

        // lm
        std::shared_ptr<fst::Fst<fst::StdArc>> lm_ = nullptr;

        string window_type = "hamming";
        int frame_length = 25;
        int frame_shift = 10;
        int n_mels = 80;
        int encoder_size = 512;
        int fsmn_layers = 16;
        int fsmn_lorder = 10;
        int fsmn_dims = 512;
        float cif_threshold = 1.0;
        float tail_alphas = 0.45;
        int asr_sample_rate = MODEL_SAMPLE_RATE;
    };

} // namespace funasr

 runtime/onnxruntime/src/precomp.h

@@ -64,6 +64,7 @@
#include "seg_dict.h"
#include "resample.h"
#include "paraformer.h"
#include "paraformer-torch.h"
#include "paraformer-online.h"
#include "offline-stream.h"
#include "tpass-stream.h"