Merge pull request #400 from alibaba-damo-academy/dev_knf

Yabin Li

2023-04-24 7ab2e5cf22bbb31808bcacf84c054c710e4e6a93

 funasr/runtime/onnxruntime/src/ct-transformer.cpp

New file
@@ -0,0 +1,188 @@
#include "precomp.h"

CTTransformer::CTTransformer(const char* sz_model_dir, int thread_num)
:env_(ORT_LOGGING_LEVEL_ERROR, ""),session_options{}
{
    session_options.SetIntraOpNumThreads(thread_num);
    session_options.SetGraphOptimizationLevel(ORT_ENABLE_ALL);
    session_options.DisableCpuMemArena();

   string strModelPath = PathAppend(sz_model_dir, PUNC_MODEL_FILE);
   string strYamlPath = PathAppend(sz_model_dir, PUNC_YAML_FILE);

    try{
#ifdef _WIN32
   std::wstring detPath = strToWstr(strModelPath);
    m_session = std::make_unique<Ort::Session>(env_, detPath.c_str(), session_options);
#else
    m_session = std::make_unique<Ort::Session>(env_, strModelPath.c_str(), session_options);
#endif
    }
    catch(exception e)
    {
        printf(e.what());
    }
    // read inputnames outputnamess
    string strName;
    GetInputName(m_session.get(), strName);
    m_strInputNames.push_back(strName.c_str());
    GetInputName(m_session.get(), strName, 1);
    m_strInputNames.push_back(strName);
    
    GetOutputName(m_session.get(), strName);
    m_strOutputNames.push_back(strName);

    for (auto& item : m_strInputNames)
        m_szInputNames.push_back(item.c_str());
    for (auto& item : m_strOutputNames)
        m_szOutputNames.push_back(item.c_str());

   m_tokenizer.OpenYaml(strYamlPath.c_str());
}

CTTransformer::~CTTransformer()
{
}

string CTTransformer::AddPunc(const char* sz_input)
{
    string strResult;
    vector<string> strOut;
    vector<int> InputData;
    m_tokenizer.Tokenize(sz_input, strOut, InputData); 

    int nTotalBatch = ceil((float)InputData.size() / TOKEN_LEN);
    int nCurBatch = -1;
    int nSentEnd = -1, nLastCommaIndex = -1;
    vector<int64_t> RemainIDs; // 
    vector<string> RemainStr; //
    vector<int> NewPunctuation; //
    vector<string> NewString; //
    vector<string> NewSentenceOut;
    vector<int> NewPuncOut;
    int nDiff = 0;
    for (size_t i = 0; i < InputData.size(); i += TOKEN_LEN)
    {
        nDiff = (i + TOKEN_LEN) < InputData.size() ? (0) : (i + TOKEN_LEN - InputData.size());
        vector<int64_t> InputIDs(InputData.begin() + i, InputData.begin() + i + TOKEN_LEN - nDiff);
        vector<string> InputStr(strOut.begin() + i, strOut.begin() + i + TOKEN_LEN - nDiff);
        InputIDs.insert(InputIDs.begin(), RemainIDs.begin(), RemainIDs.end()); // RemainIDs+InputIDs;
        InputStr.insert(InputStr.begin(), RemainStr.begin(), RemainStr.end()); // RemainStr+InputStr;

        auto Punction = Infer(InputIDs);
        nCurBatch = i / TOKEN_LEN;
        if (nCurBatch < nTotalBatch - 1) // not the last minisetence
        {
            nSentEnd = -1;
            nLastCommaIndex = -1;
            for (int nIndex = Punction.size() - 2; nIndex > 0; nIndex--)
            {
                if (m_tokenizer.Id2Punc(Punction[nIndex]) == m_tokenizer.Id2Punc(PERIOD_INDEX) || m_tokenizer.Id2Punc(Punction[nIndex]) == m_tokenizer.Id2Punc(QUESTION_INDEX))
                {
                    nSentEnd = nIndex;
                    break;
                }
                if (nLastCommaIndex < 0 && m_tokenizer.Id2Punc(Punction[nIndex]) == m_tokenizer.Id2Punc(COMMA_INDEX))
                {
                    nLastCommaIndex = nIndex;
                }
            }
            if (nSentEnd < 0 && InputStr.size() > CACHE_POP_TRIGGER_LIMIT && nLastCommaIndex > 0)
            {
                nSentEnd = nLastCommaIndex;
                Punction[nSentEnd] = PERIOD_INDEX;
            }
            RemainStr.assign(InputStr.begin() + nSentEnd + 1, InputStr.end());
            RemainIDs.assign(InputIDs.begin() + nSentEnd + 1, InputIDs.end());
            InputStr.assign(InputStr.begin(), InputStr.begin() + nSentEnd + 1);  // minit_sentence
            Punction.assign(Punction.begin(), Punction.begin() + nSentEnd + 1);
        }
        
        NewPunctuation.insert(NewPunctuation.end(), Punction.begin(), Punction.end());
        vector<string> WordWithPunc;
        for (int i = 0; i < InputStr.size(); i++)
        {
            if (i > 0 && !(InputStr[i][0] & 0x80) && (i + 1) <InputStr.size() && !(InputStr[i+1][0] & 0x80))// �м��Ӣ�ģ�
            {
                InputStr[i] = InputStr[i]+ " ";
            }
            WordWithPunc.push_back(InputStr[i]);

            if (Punction[i] != NOTPUNC_INDEX) // �»���
            {
                WordWithPunc.push_back(m_tokenizer.Id2Punc(Punction[i]));
            }
        }

        NewString.insert(NewString.end(), WordWithPunc.begin(), WordWithPunc.end()); // new_mini_sentence += "".join(words_with_punc)
        NewSentenceOut = NewString;
        NewPuncOut = NewPunctuation;
        // last mini sentence
        if(nCurBatch == nTotalBatch - 1)
        {
            if (NewString[NewString.size() - 1] == m_tokenizer.Id2Punc(COMMA_INDEX) || NewString[NewString.size() - 1] == m_tokenizer.Id2Punc(DUN_INDEX))
            {
                NewSentenceOut.assign(NewString.begin(), NewString.end() - 1);
                NewSentenceOut.push_back(m_tokenizer.Id2Punc(PERIOD_INDEX));
                NewPuncOut.assign(NewPunctuation.begin(), NewPunctuation.end() - 1);
                NewPuncOut.push_back(PERIOD_INDEX);
            }
            else if (NewString[NewString.size() - 1] == m_tokenizer.Id2Punc(PERIOD_INDEX) && NewString[NewString.size() - 1] == m_tokenizer.Id2Punc(QUESTION_INDEX))
            {
                NewSentenceOut = NewString;
                NewSentenceOut.push_back(m_tokenizer.Id2Punc(PERIOD_INDEX));
                NewPuncOut = NewPunctuation;
                NewPuncOut.push_back(PERIOD_INDEX);
            }
        }
    }
    for (auto& item : NewSentenceOut)
        strResult += item;
    return strResult;
}

vector<int> CTTransformer::Infer(vector<int64_t> input_data)
{
    Ort::MemoryInfo m_memoryInfo = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
    vector<int> punction;
    std::array<int64_t, 2> input_shape_{ 1, (int64_t)input_data.size()};
    Ort::Value onnx_input = Ort::Value::CreateTensor<int64_t>(m_memoryInfo,
        input_data.data(),
        input_data.size(),
        input_shape_.data(),
        input_shape_.size());

    std::array<int32_t,1> text_lengths{ (int32_t)input_data.size() };
    std::array<int64_t,1> text_lengths_dim{ 1 };
    Ort::Value onnx_text_lengths = Ort::Value::CreateTensor(
        m_memoryInfo,
        text_lengths.data(),
        text_lengths.size() * sizeof(int32_t),
        text_lengths_dim.data(),
        text_lengths_dim.size(), ONNX_TENSOR_ELEMENT_DATA_TYPE_INT32);
    std::vector<Ort::Value> input_onnx;
    input_onnx.emplace_back(std::move(onnx_input));
    input_onnx.emplace_back(std::move(onnx_text_lengths));
        
    try {
        auto outputTensor = m_session->Run(Ort::RunOptions{nullptr}, m_szInputNames.data(), input_onnx.data(), m_szInputNames.size(), m_szOutputNames.data(), m_szOutputNames.size());
        std::vector<int64_t> outputShape = outputTensor[0].GetTensorTypeAndShapeInfo().GetShape();

        int64_t outputCount = std::accumulate(outputShape.begin(), outputShape.end(), 1, std::multiplies<int64_t>());
        float * floatData = outputTensor[0].GetTensorMutableData<float>();

        for (int i = 0; i < outputCount; i += CANDIDATE_NUM)
        {
            int index = Argmax(floatData + i, floatData + i + CANDIDATE_NUM-1);
            punction.push_back(index);
        }
    }
    catch (std::exception const &e)
    {
        printf(e.what());
    }
    return punction;
}