| | |
| | | */ |
| | | |
| | | #include "precomp.h" |
| | | #include "paraformer.h" |
| | | #include "encode_converter.h" |
| | | #include <cstddef> |
| | | |
| | | using namespace std; |
| | | |
| | | namespace funasr { |
| | | |
| | | Paraformer::Paraformer() |
| | | :env_(ORT_LOGGING_LEVEL_ERROR, "paraformer"),session_options_{}{ |
| | | :use_hotword(false), |
| | | env_(ORT_LOGGING_LEVEL_ERROR, "paraformer"),session_options_{}, |
| | | hw_env_(ORT_LOGGING_LEVEL_ERROR, "paraformer_hw"),hw_session_options{} { |
| | | } |
| | | |
| | | // offline |
| | |
| | | m_strInputNames.push_back(strName.c_str()); |
| | | GetInputName(m_session_.get(), strName,1); |
| | | m_strInputNames.push_back(strName); |
| | | if (use_hotword) { |
| | | GetInputName(m_session_.get(), strName, 2); |
| | | m_strInputNames.push_back(strName); |
| | | } |
| | | |
| | | size_t numOutputNodes = m_session_->GetOutputCount(); |
| | | for(int index=0; index<numOutputNodes; index++){ |
| | |
| | | } |
| | | } |
| | | |
| | | void Paraformer::InitHwCompiler(const std::string &hw_model, int thread_num) { |
| | | hw_session_options.SetIntraOpNumThreads(thread_num); |
| | | hw_session_options.SetGraphOptimizationLevel(ORT_ENABLE_ALL); |
| | | // DisableCpuMemArena can improve performance |
| | | hw_session_options.DisableCpuMemArena(); |
| | | |
| | | try { |
| | | hw_m_session = std::make_unique<Ort::Session>(hw_env_, hw_model.c_str(), hw_session_options); |
| | | LOG(INFO) << "Successfully load model from " << hw_model; |
| | | } catch (std::exception const &e) { |
| | | LOG(ERROR) << "Error when load hw compiler onnx model: " << e.what(); |
| | | exit(0); |
| | | } |
| | | |
| | | string strName; |
| | | GetInputName(hw_m_session.get(), strName); |
| | | hw_m_strInputNames.push_back(strName.c_str()); |
| | | //GetInputName(hw_m_session.get(), strName,1); |
| | | //hw_m_strInputNames.push_back(strName); |
| | | |
| | | GetOutputName(hw_m_session.get(), strName); |
| | | hw_m_strOutputNames.push_back(strName); |
| | | |
| | | for (auto& item : hw_m_strInputNames) |
| | | hw_m_szInputNames.push_back(item.c_str()); |
| | | for (auto& item : hw_m_strOutputNames) |
| | | hw_m_szOutputNames.push_back(item.c_str()); |
| | | // if init hotword compiler is called, this is a hotword paraformer model |
| | | use_hotword = true; |
| | | } |
| | | |
| | | void Paraformer::InitSegDict(const std::string &seg_dict_model) { |
| | | seg_dict = new SegDict(seg_dict_model.c_str()); |
| | | } |
| | | |
| | | Paraformer::~Paraformer() |
| | | { |
| | | if(vocab) |
| | | delete vocab; |
| | | if(seg_dict) |
| | | delete seg_dict; |
| | | } |
| | | |
| | | void Paraformer::Reset() |
| | |
| | | int32_t feature_dim = fbank_opts_.mel_opts.num_bins; |
| | | vector<float> features(frames * feature_dim); |
| | | float *p = features.data(); |
| | | //std::cout << "samples " << len << std::endl; |
| | | //std::cout << "fbank frames " << frames << std::endl; |
| | | //std::cout << "fbank dim " << feature_dim << std::endl; |
| | | //std::cout << "feature size " << features.size() << std::endl; |
| | | |
| | | for (int32_t i = 0; i != frames; ++i) { |
| | | const float *f = fbank_.GetFrame(i); |
| | |
| | | } |
| | | } |
| | | |
| | | string Paraformer::Forward(float* din, int len, bool input_finished) |
| | | string Paraformer::Forward(float* din, int len, bool input_finished, const std::vector<std::vector<float>> &hw_emb) |
| | | { |
| | | |
| | | int32_t in_feat_dim = fbank_opts_.mel_opts.num_bins; |
| | |
| | | |
| | | int32_t feat_dim = lfr_m*in_feat_dim; |
| | | int32_t num_frames = wav_feats.size() / feat_dim; |
| | | //std::cout << "feat in: " << num_frames << " " << feat_dim << std::endl; |
| | | |
| | | #ifdef _WIN_X86 |
| | | Ort::MemoryInfo m_memoryInfo = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU); |
| | |
| | | paraformer_length.emplace_back(num_frames); |
| | | Ort::Value onnx_feats_len = Ort::Value::CreateTensor<int32_t>( |
| | | m_memoryInfo, paraformer_length.data(), paraformer_length.size(), paraformer_length_shape, 1); |
| | | |
| | | |
| | | std::vector<Ort::Value> input_onnx; |
| | | input_onnx.emplace_back(std::move(onnx_feats)); |
| | | input_onnx.emplace_back(std::move(onnx_feats_len)); |
| | | |
| | | std::vector<float> embedding; |
| | | try{ |
| | | if (use_hotword) { |
| | | if(hw_emb.size()<=0){ |
| | | LOG(ERROR) << "hw_emb is null"; |
| | | return ""; |
| | | } |
| | | //PrintMat(hw_emb, "input_clas_emb"); |
| | | const int64_t hotword_shape[3] = {1, hw_emb.size(), hw_emb[0].size()}; |
| | | embedding.reserve(hw_emb.size() * hw_emb[0].size()); |
| | | for (auto item : hw_emb) { |
| | | embedding.insert(embedding.end(), item.begin(), item.end()); |
| | | } |
| | | //LOG(INFO) << "hotword shape " << hotword_shape[0] << " " << hotword_shape[1] << " " << hotword_shape[2] << " size " << embedding.size(); |
| | | Ort::Value onnx_hw_emb = Ort::Value::CreateTensor<float>( |
| | | m_memoryInfo, embedding.data(), embedding.size(), hotword_shape, 3); |
| | | |
| | | input_onnx.emplace_back(std::move(onnx_hw_emb)); |
| | | } |
| | | }catch (std::exception const &e) |
| | | { |
| | | LOG(ERROR)<<e.what(); |
| | | return ""; |
| | | } |
| | | |
| | | string result; |
| | | try { |
| | | auto outputTensor = m_session_->Run(Ort::RunOptions{nullptr}, m_szInputNames.data(), input_onnx.data(), input_onnx.size(), m_szOutputNames.data(), m_szOutputNames.size()); |
| | | std::vector<int64_t> outputShape = outputTensor[0].GetTensorTypeAndShapeInfo().GetShape(); |
| | | //LOG(INFO) << "paraformer out shape " << outputShape[0] << " " << outputShape[1] << " " << outputShape[2]; |
| | | |
| | | int64_t outputCount = std::accumulate(outputShape.begin(), outputShape.end(), 1, std::multiplies<int64_t>()); |
| | | float* floatData = outputTensor[0].GetTensorMutableData<float>(); |
| | |
| | | }else{ |
| | | result = GreedySearch(floatData, *encoder_out_lens, outputShape[2]); |
| | | } |
| | | // int pos = 0; |
| | | // std::vector<std::vector<float>> logits; |
| | | // for (int j = 0; j < outputShape[1]; j++) |
| | | // { |
| | | // std::vector<float> vec_token; |
| | | // vec_token.insert(vec_token.begin(), floatData + pos, floatData + pos + outputShape[2]); |
| | | // logits.push_back(vec_token); |
| | | // pos += outputShape[2]; |
| | | // } |
| | | // //PrintMat(logits, "logits_out"); |
| | | // result = GreedySearch(floatData, *encoder_out_lens, outputShape[2]); |
| | | } |
| | | catch (std::exception const &e) |
| | | { |
| | |
| | | return result; |
| | | } |
| | | |
| | | |
| | | std::vector<std::vector<float>> Paraformer::CompileHotwordEmbedding(std::string &hotwords) { |
| | | int embedding_dim = encoder_size; |
| | | std::vector<std::vector<float>> hw_emb; |
| | | if (!use_hotword) { |
| | | std::vector<float> vec(embedding_dim, 0); |
| | | hw_emb.push_back(vec); |
| | | return hw_emb; |
| | | } |
| | | int max_hotword_len = 10; |
| | | std::vector<int32_t> hotword_matrix; |
| | | std::vector<int32_t> lengths; |
| | | int hotword_size = 1; |
| | | if (!hotwords.empty()) { |
| | | std::vector<std::string> hotword_array = split(hotwords, ' '); |
| | | hotword_size = hotword_array.size() + 1; |
| | | hotword_matrix.reserve(hotword_size * max_hotword_len); |
| | | for (auto hotword : hotword_array) { |
| | | std::vector<std::string> chars; |
| | | if (EncodeConverter::IsAllChineseCharactor((const U8CHAR_T*)hotword.c_str(), hotword.size())) { |
| | | KeepChineseCharacterAndSplit(hotword, chars); |
| | | } else { |
| | | // for english |
| | | std::vector<std::string> words = split(hotword, ' '); |
| | | for (auto word : words) { |
| | | std::vector<string> tokens = seg_dict->GetTokensByWord(word); |
| | | chars.insert(chars.end(), tokens.begin(), tokens.end()); |
| | | } |
| | | } |
| | | std::vector<int32_t> hw_vector(max_hotword_len, 0); |
| | | int vector_len = std::min(max_hotword_len, (int)chars.size()); |
| | | for (int i=0; i<chars.size(); i++) { |
| | | std::cout << chars[i] << " "; |
| | | hw_vector[i] = vocab->GetIdByToken(chars[i]); |
| | | } |
| | | std::cout << std::endl; |
| | | lengths.push_back(vector_len); |
| | | hotword_matrix.insert(hotword_matrix.end(), hw_vector.begin(), hw_vector.end()); |
| | | } |
| | | } |
| | | std::vector<int32_t> blank_vec(max_hotword_len, 0); |
| | | blank_vec[0] = 1; |
| | | hotword_matrix.insert(hotword_matrix.end(), blank_vec.begin(), blank_vec.end()); |
| | | lengths.push_back(1); |
| | | |
| | | #ifdef _WIN_X86 |
| | | Ort::MemoryInfo m_memoryInfo = Ort::MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU); |
| | | #else |
| | | Ort::MemoryInfo m_memoryInfo = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault); |
| | | #endif |
| | | |
| | | const int64_t input_shape_[2] = {hotword_size, max_hotword_len}; |
| | | Ort::Value onnx_hotword = Ort::Value::CreateTensor<int32_t>(m_memoryInfo, |
| | | (int32_t*)hotword_matrix.data(), |
| | | hotword_size * max_hotword_len, |
| | | input_shape_, |
| | | 2); |
| | | LOG(INFO) << "clas shape " << hotword_size << " " << max_hotword_len << std::endl; |
| | | |
| | | std::vector<Ort::Value> input_onnx; |
| | | input_onnx.emplace_back(std::move(onnx_hotword)); |
| | | |
| | | std::vector<std::vector<float>> result; |
| | | try { |
| | | auto outputTensor = hw_m_session->Run(Ort::RunOptions{nullptr}, hw_m_szInputNames.data(), input_onnx.data(), input_onnx.size(), hw_m_szOutputNames.data(), hw_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>(); // shape [max_hotword_len, hotword_size, dim] |
| | | // get embedding by real hotword length |
| | | assert(outputShape[0] == max_hotword_len); |
| | | assert(outputShape[1] == hotword_size); |
| | | embedding_dim = outputShape[2]; |
| | | |
| | | for (int j = 0; j < hotword_size; j++) |
| | | { |
| | | int start_pos = hotword_size * (lengths[j] - 1) * embedding_dim + j * embedding_dim; |
| | | std::vector<float> embedding; |
| | | embedding.insert(embedding.begin(), floatData + start_pos, floatData + start_pos + embedding_dim); |
| | | result.push_back(embedding); |
| | | } |
| | | } |
| | | catch (std::exception const &e) |
| | | { |
| | | LOG(ERROR)<<e.what(); |
| | | } |
| | | //PrintMat(result, "clas_embedding_output"); |
| | | return result; |
| | | } |
| | | |
| | | string Paraformer::Rescoring() |
| | | { |
| | | LOG(ERROR)<<"Not Imp!!!!!!"; |