From ec5e99be3ae6dd0d3811950f155b01e144431452 Mon Sep 17 00:00:00 2001 From: Richard Burton Date: Wed, 5 Oct 2022 11:00:37 +0100 Subject: MLECO-3164: Additional refactoring of KWS API Part 1 * Add KwsClassifier * KwsPostProcess can now be told to average results * Averaging is handlded by KwsClassifier * Current sliding window index is now an argument of DoPreProcess Change-Id: I07626da595ad1cbd982e8366f0d1bb56d1040459 --- .../application/api/common/include/Classifier.hpp | 10 +- source/application/api/common/include/Model.hpp | 12 +- source/application/api/common/source/Classifier.cc | 22 ++-- .../api/use_case/ad/include/AdMelSpectrogram.hpp | 4 +- source/application/api/use_case/kws/CMakeLists.txt | 3 +- .../api/use_case/kws/include/KwsClassifier.hpp | 66 ++++++++++ .../api/use_case/kws/include/KwsProcessing.hpp | 19 ++- .../api/use_case/kws/include/KwsResult.hpp | 4 +- .../api/use_case/kws/src/KwsClassifier.cc | 142 +++++++++++++++++++++ .../api/use_case/kws/src/KwsProcessing.cc | 19 +-- .../main/include/UseCaseCommonUtils.hpp | 1 - source/use_case/kws/src/MainLoop.cc | 9 +- source/use_case/kws/src/UseCaseHandler.cc | 9 +- source/use_case/kws_asr/src/MainLoop.cc | 8 +- source/use_case/kws_asr/src/UseCaseHandler.cc | 5 +- 15 files changed, 263 insertions(+), 70 deletions(-) create mode 100644 source/application/api/use_case/kws/include/KwsClassifier.hpp create mode 100644 source/application/api/use_case/kws/src/KwsClassifier.cc (limited to 'source') diff --git a/source/application/api/common/include/Classifier.hpp b/source/application/api/common/include/Classifier.hpp index d641c22..e4eab01 100644 --- a/source/application/api/common/include/Classifier.hpp +++ b/source/application/api/common/include/Classifier.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. All rights reserved. + * Copyright (c) 2021-2022 Arm Limited. All rights reserved. * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the "License"); @@ -34,6 +34,8 @@ namespace app { /** @brief Constructor. */ Classifier() = default; + virtual ~Classifier() = default; + /** * @brief Gets the top N classification results from the * output vector. @@ -41,8 +43,8 @@ namespace app { * @param[out] vecResults A vector of classification results. * populated by this function. * @param[in] labels Labels vector to match classified classes. - * @param[in] topNCount Number of top classifications to pick. Default is 1. - * @param[in] useSoftmax Whether Softmax normalisation should be applied to output. Default is false. + * @param[in] topNCount Number of top classifications to pick. + * @param[in] useSoftmax Whether Softmax normalisation should be applied to output. * @return true if successful, false otherwise. **/ @@ -65,7 +67,7 @@ namespace app { std::vector& vecResults, const std::vector & labels); - private: + protected: /** * @brief Utility function that gets the top N classification results from the * output vector. diff --git a/source/application/api/common/include/Model.hpp b/source/application/api/common/include/Model.hpp index 70c6245..4892757 100644 --- a/source/application/api/common/include/Model.hpp +++ b/source/application/api/common/include/Model.hpp @@ -137,13 +137,13 @@ namespace app { const tflite::Model* m_pModel{nullptr}; /* Tflite model pointer. */ tflite::MicroInterpreter* m_pInterpreter{nullptr}; /* Tflite interpreter. */ tflite::MicroAllocator* m_pAllocator{nullptr}; /* Tflite micro allocator. */ - bool m_inited{false}; /* Indicates whether this object has been initialised. */ - const uint8_t* m_modelAddr{nullptr}; /* Model address */ - uint32_t m_modelSize{0}; /* Model size */ + bool m_inited{false}; /* Indicates whether this object has been initialised. */ + const uint8_t* m_modelAddr{nullptr}; /* Model address */ + uint32_t m_modelSize{0}; /* Model size */ - std::vector m_input{}; /* Model's input tensor pointers. */ - std::vector m_output{}; /* Model's output tensor pointers. */ - TfLiteType m_type{kTfLiteNoType}; /* Model's data type. */ + std::vector m_input{}; /* Model's input tensor pointers. */ + std::vector m_output{}; /* Model's output tensor pointers. */ + TfLiteType m_type{kTfLiteNoType}; /* Model's data type. */ }; } /* namespace app */ diff --git a/source/application/api/common/source/Classifier.cc b/source/application/api/common/source/Classifier.cc index 6fabebe..1b5fc64 100644 --- a/source/application/api/common/source/Classifier.cc +++ b/source/application/api/common/source/Classifier.cc @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. All rights reserved. + * Copyright (c) 2021-2022 Arm Limited. All rights reserved. * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the "License"); @@ -31,10 +31,9 @@ namespace arm { namespace app { void Classifier::SetVectorResults(std::set>& topNSet, - std::vector& vecResults, - const std::vector & labels) + std::vector& vecResults, + const std::vector & labels) { - /* Reset the iterator to the largest element - use reverse iterator. */ auto topNIter = topNSet.rbegin(); @@ -46,11 +45,9 @@ namespace app { } bool Classifier::GetTopNResults(const std::vector& tensor, - std::vector& vecResults, - uint32_t topNCount, - const std::vector & labels) + std::vector& vecResults, + uint32_t topNCount, const std::vector & labels) { - std::set> sortedSet; /* NOTE: inputVec's size verification against labels should be @@ -80,12 +77,9 @@ namespace app { return true; } - bool Classifier::GetClassificationResults( - TfLiteTensor* outputTensor, - std::vector& vecResults, - const std::vector & labels, - uint32_t topNCount, - bool useSoftmax) + bool Classifier::GetClassificationResults(TfLiteTensor* outputTensor, + std::vector& vecResults, const std::vector & labels, + uint32_t topNCount, bool useSoftmax) { if (outputTensor == nullptr) { printf_err("Output vector is null pointer.\n"); diff --git a/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp b/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp index 05c5bfc..b8a9dfc 100644 --- a/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp +++ b/source/application/api/use_case/ad/include/AdMelSpectrogram.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. All rights reserved. + * Copyright (c) 2021-2022 Arm Limited. All rights reserved. * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the "License"); @@ -41,7 +41,7 @@ namespace audio { {} AdMelSpectrogram() = delete; - ~AdMelSpectrogram() = default; + virtual ~AdMelSpectrogram() = default; protected: diff --git a/source/application/api/use_case/kws/CMakeLists.txt b/source/application/api/use_case/kws/CMakeLists.txt index 517a35a..ea0761c 100644 --- a/source/application/api/use_case/kws/CMakeLists.txt +++ b/source/application/api/use_case/kws/CMakeLists.txt @@ -27,7 +27,8 @@ project(${KWS_API_TARGET} # Create static library add_library(${KWS_API_TARGET} STATIC src/KwsProcessing.cc - src/MicroNetKwsModel.cc) + src/MicroNetKwsModel.cc + src/KwsClassifier.cc) target_include_directories(${KWS_API_TARGET} PUBLIC include) diff --git a/source/application/api/use_case/kws/include/KwsClassifier.hpp b/source/application/api/use_case/kws/include/KwsClassifier.hpp new file mode 100644 index 0000000..d050e85 --- /dev/null +++ b/source/application/api/use_case/kws/include/KwsClassifier.hpp @@ -0,0 +1,66 @@ +/* + * Copyright (c) 2022 Arm Limited. All rights reserved. + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#ifndef KWS_CLASSIFIER_HPP +#define KWS_CLASSIFIER_HPP + +#include "ClassificationResult.hpp" +#include "TensorFlowLiteMicro.hpp" +#include "Classifier.hpp" + +#include + +namespace arm { +namespace app { + + /** + * @brief KWS Classifier - a helper class to get certain number of top + * results from the output vector from a classification NN. + * Allows for averaging of previous results. + **/ + class KwsClassifier : public Classifier { + public: + + /** + * @brief Gets the top N classification results from the + * output vector. + * @param[in] outputTensor Inference output tensor from an NN model. + * @param[out] vecResults A vector of classification results. + * populated by this function. + * @param[in] labels Labels vector to match classified classes. + * @param[in] topNCount Number of top classifications to pick. Default is 1. + * @param[in] useSoftmax Whether Softmax normalisation should be applied to output. Default is false. + * @param[in/out] resultHistory History of previous classification results to be updated. + * @return true if successful, false otherwise. + **/ + using Classifier::GetClassificationResults; /* We are overloading not overriding. */ + bool GetClassificationResults(TfLiteTensor* outputTensor, std::vector& vecResults, + const std::vector & labels, uint32_t topNCount, + bool use_softmax, std::vector>& resultHistory); + + /** + * @brief Average the given history of results. + * @param[in] resultHistory The history of results to take on average of. + * @param[out] averageResult The calculated average. + **/ + static void AveragResults(const std::vector>& resultHistory, + std::vector& averageResult); + }; + +} /* namespace app */ +} /* namespace arm */ + +#endif /* KWS_CLASSIFIER_HPP */ diff --git a/source/application/api/use_case/kws/include/KwsProcessing.hpp b/source/application/api/use_case/kws/include/KwsProcessing.hpp index 0ede425..e2d3ff9 100644 --- a/source/application/api/use_case/kws/include/KwsProcessing.hpp +++ b/source/application/api/use_case/kws/include/KwsProcessing.hpp @@ -19,7 +19,7 @@ #include "AudioUtils.hpp" #include "BaseProcessing.hpp" -#include "Classifier.hpp" +#include "KwsClassifier.hpp" #include "MicroNetKwsMfcc.hpp" #include @@ -55,9 +55,8 @@ namespace app { * @param[in] inputSize Size of the input data. * @return true if successful, false otherwise. **/ - bool DoPreProcess(const void* input, size_t inputSize) override; + bool DoPreProcess(const void* input, size_t inferenceIndex = 0) override; - size_t m_audioWindowIndex = 0; /* Index of audio slider, used when caching features in longer clips. */ size_t m_audioDataWindowSize; /* Amount of audio needed for 1 inference. */ size_t m_audioDataStride; /* Amount of audio to stride across if doing >1 inference in longer clips. */ @@ -106,11 +105,11 @@ namespace app { class KwsPostProcess : public BasePostProcess { private: - TfLiteTensor* m_outputTensor; /* Model output tensor. */ - Classifier& m_kwsClassifier; /* KWS Classifier object. */ - const std::vector& m_labels; /* KWS Labels. */ - std::vector& m_results; /* Results vector for a single inference. */ - + TfLiteTensor* m_outputTensor; /* Model output tensor. */ + KwsClassifier& m_kwsClassifier; /* KWS Classifier object. */ + const std::vector& m_labels; /* KWS Labels. */ + std::vector& m_results; /* Results vector for a single inference. */ + std::vector> m_resultHistory; /* Store previous results so they can be averaged. */ public: /** * @brief Constructor @@ -119,9 +118,9 @@ namespace app { * @param[in] labels Vector of string labels to identify each output of the model. * @param[in/out] results Vector of classification results to store decoded outputs. **/ - KwsPostProcess(TfLiteTensor* outputTensor, Classifier& classifier, + KwsPostProcess(TfLiteTensor* outputTensor, KwsClassifier& classifier, const std::vector& labels, - std::vector& results); + std::vector& results, size_t averagingWindowLen = 1); /** * @brief Should perform post-processing of the result of inference then diff --git a/source/application/api/use_case/kws/include/KwsResult.hpp b/source/application/api/use_case/kws/include/KwsResult.hpp index 38f32b4..e0bb868 100644 --- a/source/application/api/use_case/kws/include/KwsResult.hpp +++ b/source/application/api/use_case/kws/include/KwsResult.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. All rights reserved. + * Copyright (c) 2021-2022 Arm Limited. All rights reserved. * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the "License"); @@ -47,7 +47,7 @@ namespace kws { this->m_inferenceNumber = inferenceIdx; this->m_resultVec = ResultVec(); - for (auto & i : resultVec) { + for (auto& i : resultVec) { if (i.m_normalisedVal >= this->m_threshold) { this->m_resultVec.emplace_back(i); } diff --git a/source/application/api/use_case/kws/src/KwsClassifier.cc b/source/application/api/use_case/kws/src/KwsClassifier.cc new file mode 100644 index 0000000..fe409b1 --- /dev/null +++ b/source/application/api/use_case/kws/src/KwsClassifier.cc @@ -0,0 +1,142 @@ +/* + * Copyright (c) 2022 Arm Limited. All rights reserved. + * SPDX-License-Identifier: Apache-2.0 + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +#include "KwsClassifier.hpp" + +#include "TensorFlowLiteMicro.hpp" +#include "PlatformMath.hpp" +#include "log_macros.h" +#include "../include/KwsClassifier.hpp" + + +#include +#include +#include +#include +#include +#include + + +namespace arm { +namespace app { + + bool KwsClassifier::GetClassificationResults(TfLiteTensor* outputTensor, + std::vector& vecResults, const std::vector & labels, + uint32_t topNCount, bool useSoftmax, std::vector>& resultHistory) + { + if (outputTensor == nullptr) { + printf_err("Output vector is null pointer.\n"); + return false; + } + + uint32_t totalOutputSize = 1; + for (int inputDim = 0; inputDim < outputTensor->dims->size; inputDim++) { + totalOutputSize *= outputTensor->dims->data[inputDim]; + } + + /* Sanity checks. */ + if (totalOutputSize < topNCount) { + printf_err("Output vector is smaller than %" PRIu32 "\n", topNCount); + return false; + } else if (totalOutputSize != labels.size()) { + printf_err("Output size doesn't match the labels' size\n"); + return false; + } else if (topNCount == 0) { + printf_err("Top N results cannot be zero\n"); + return false; + } + + bool resultState; + vecResults.clear(); + + /* De-Quantize Output Tensor */ + QuantParams quantParams = GetTensorQuantParams(outputTensor); + + /* Floating point tensor data to be populated + * NOTE: The assumption here is that the output tensor size isn't too + * big and therefore, there's neglibible impact on heap usage. */ + std::vector resultData(totalOutputSize); + resultData.resize(totalOutputSize); + + /* Populate the floating point buffer */ + switch (outputTensor->type) { + case kTfLiteUInt8: { + uint8_t* tensor_buffer = tflite::GetTensorData(outputTensor); + for (size_t i = 0; i < totalOutputSize; ++i) { + resultData[i] = quantParams.scale * + (static_cast(tensor_buffer[i]) - quantParams.offset); + } + break; + } + case kTfLiteInt8: { + int8_t* tensor_buffer = tflite::GetTensorData(outputTensor); + for (size_t i = 0; i < totalOutputSize; ++i) { + resultData[i] = quantParams.scale * + (static_cast(tensor_buffer[i]) - quantParams.offset); + } + break; + } + case kTfLiteFloat32: { + float* tensor_buffer = tflite::GetTensorData(outputTensor); + for (size_t i = 0; i < totalOutputSize; ++i) { + resultData[i] = tensor_buffer[i]; + } + break; + } + default: + printf_err("Tensor type %s not supported by classifier\n", + TfLiteTypeGetName(outputTensor->type)); + return false; + } + + if (useSoftmax) { + math::MathUtils::SoftmaxF32(resultData); + } + + /* If keeping track of recent results, update and take an average. */ + if (resultHistory.size() > 1) { + std::rotate(resultHistory.begin(), resultHistory.begin() + 1, resultHistory.end()); + resultHistory.back() = resultData; + AveragResults(resultHistory, resultData); + } + + /* Get the top N results. */ + resultState = GetTopNResults(resultData, vecResults, topNCount, labels); + + if (!resultState) { + printf_err("Failed to get top N results set\n"); + return false; + } + + return true; + } + + void app::KwsClassifier::AveragResults(const std::vector>& resultHistory, + std::vector& averageResult) + { + /* Compute averages of each class across the window length. */ + float sum; + for (size_t j = 0; j < averageResult.size(); j++) { + sum = 0; + for (size_t i = 0; i < resultHistory.size(); i++) { + sum += resultHistory[i][j]; + } + averageResult[j] = (sum / resultHistory.size()); + } + } + +} /* namespace app */ +} /* namespace arm */ \ No newline at end of file diff --git a/source/application/api/use_case/kws/src/KwsProcessing.cc b/source/application/api/use_case/kws/src/KwsProcessing.cc index 2d5c085..843ac58 100644 --- a/source/application/api/use_case/kws/src/KwsProcessing.cc +++ b/source/application/api/use_case/kws/src/KwsProcessing.cc @@ -66,9 +66,8 @@ namespace app { } } - bool KwsPreProcess::DoPreProcess(const void* data, size_t inputSize) + bool KwsPreProcess::DoPreProcess(const void* data, size_t inferenceIndex) { - UNUSED(inputSize); if (data == nullptr) { printf_err("Data pointer is null"); } @@ -77,8 +76,8 @@ namespace app { auto input = static_cast(data); this->m_mfccSlidingWindow.Reset(input); - /* Cache is only usable if we have more than 1 inference in an audio clip. */ - bool useCache = this->m_audioWindowIndex > 0 && this->m_numReusedMfccVectors > 0; + /* Cache is only usable if we have more than 1 inference to do and it's not the first inference. */ + bool useCache = inferenceIndex > 0 && this->m_numReusedMfccVectors > 0; /* Use a sliding window to calculate MFCC features frame by frame. */ while (this->m_mfccSlidingWindow.HasNext()) { @@ -163,7 +162,7 @@ namespace app { TfLiteQuantization quant = inputTensor->quantization; if (kTfLiteAffineQuantization == quant.type) { - auto *quantParams = (TfLiteAffineQuantization *) quant.params; + auto* quantParams = (TfLiteAffineQuantization*) quant.params; const float quantScale = quantParams->scale->data[0]; const int quantOffset = quantParams->zero_point->data[0]; @@ -191,20 +190,22 @@ namespace app { return mfccFeatureCalc; } - KwsPostProcess::KwsPostProcess(TfLiteTensor* outputTensor, Classifier& classifier, + KwsPostProcess::KwsPostProcess(TfLiteTensor* outputTensor, KwsClassifier& classifier, const std::vector& labels, - std::vector& results) + std::vector& results, size_t averagingWindowLen) :m_outputTensor{outputTensor}, m_kwsClassifier{classifier}, m_labels{labels}, m_results{results} - {} + { + this->m_resultHistory = {averagingWindowLen, std::vector(labels.size())}; + } bool KwsPostProcess::DoPostProcess() { return this->m_kwsClassifier.GetClassificationResults( this->m_outputTensor, this->m_results, - this->m_labels, 1, true); + this->m_labels, 1, true, this->m_resultHistory); } } /* namespace app */ diff --git a/source/application/main/include/UseCaseCommonUtils.hpp b/source/application/main/include/UseCaseCommonUtils.hpp index b0f2e7a..9b6d550 100644 --- a/source/application/main/include/UseCaseCommonUtils.hpp +++ b/source/application/main/include/UseCaseCommonUtils.hpp @@ -24,7 +24,6 @@ #include "UseCaseHandler.hpp" /* Handlers for different user options. */ #include "Classifier.hpp" /* Classifier. */ #include "InputFiles.hpp" -#include "BufAttributes.hpp" /* Buffer attributes */ void DisplayCommonMenu(); diff --git a/source/use_case/kws/src/MainLoop.cc b/source/use_case/kws/src/MainLoop.cc index e0518f2..2489df8 100644 --- a/source/use_case/kws/src/MainLoop.cc +++ b/source/use_case/kws/src/MainLoop.cc @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. All rights reserved. + * Copyright (c) 2021-2022 Arm Limited. All rights reserved. * SPDX-License-Identifier: Apache-2.0 * * Licensed under the Apache License, Version 2.0 (the "License"); @@ -15,7 +15,7 @@ * limitations under the License. */ #include "InputFiles.hpp" /* For input audio clips. */ -#include "Classifier.hpp" /* Classifier. */ +#include "KwsClassifier.hpp" /* Classifier. */ #include "MicroNetKwsModel.hpp" /* Model class for running inference. */ #include "hal.h" /* Brings in platform definitions. */ #include "Labels.hpp" /* For label strings. */ @@ -34,7 +34,6 @@ namespace app { } /* namespace app */ } /* namespace arm */ -using KwsClassifier = arm::app::Classifier; enum opcodes { @@ -83,8 +82,8 @@ void main_loop() caseContext.Set("frameStride", arm::app::kws::g_FrameStride); caseContext.Set("scoreThreshold", arm::app::kws::g_ScoreThreshold); /* Normalised score threshold. */ - KwsClassifier classifier; /* classifier wrapper object. */ - caseContext.Set("classifier", classifier); + arm::app::KwsClassifier classifier; /* classifier wrapper object. */ + caseContext.Set("classifier", classifier); std::vector labels; GetLabelsVector(labels); diff --git a/source/use_case/kws/src/UseCaseHandler.cc b/source/use_case/kws/src/UseCaseHandler.cc index 61c6eb6..d61ba9d 100644 --- a/source/use_case/kws/src/UseCaseHandler.cc +++ b/source/use_case/kws/src/UseCaseHandler.cc @@ -17,7 +17,7 @@ #include "UseCaseHandler.hpp" #include "InputFiles.hpp" -#include "Classifier.hpp" +#include "KwsClassifier.hpp" #include "MicroNetKwsModel.hpp" #include "hal.h" #include "AudioUtils.hpp" @@ -29,8 +29,6 @@ #include -using KwsClassifier = arm::app::Classifier; - namespace arm { namespace app { @@ -124,14 +122,11 @@ namespace app { while (audioDataSlider.HasNext()) { const int16_t* inferenceWindow = audioDataSlider.Next(); - /* The first window does not have cache ready. */ - preProcess.m_audioWindowIndex = audioDataSlider.Index(); - info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, audioDataSlider.TotalStrides() + 1); /* Run the pre-processing, inference and post-processing. */ - if (!preProcess.DoPreProcess(inferenceWindow, audio::MicroNetKwsMFCC::ms_defaultSamplingFreq)) { + if (!preProcess.DoPreProcess(inferenceWindow, audioDataSlider.Index())) { printf_err("Pre-processing failed."); return false; } diff --git a/source/use_case/kws_asr/src/MainLoop.cc b/source/use_case/kws_asr/src/MainLoop.cc index 0638ecd..a4f7db9 100644 --- a/source/use_case/kws_asr/src/MainLoop.cc +++ b/source/use_case/kws_asr/src/MainLoop.cc @@ -17,7 +17,7 @@ #include "InputFiles.hpp" /* For input images. */ #include "Labels_micronetkws.hpp" /* For MicroNetKws label strings. */ #include "Labels_wav2letter.hpp" /* For Wav2Letter label strings. */ -#include "Classifier.hpp" /* KWS classifier. */ +#include "KwsClassifier.hpp" /* KWS classifier. */ #include "AsrClassifier.hpp" /* ASR classifier. */ #include "MicroNetKwsModel.hpp" /* KWS model class for running inference. */ #include "Wav2LetterModel.hpp" /* ASR model class for running inference. */ @@ -42,8 +42,6 @@ namespace app { } /* namespace app */ } /* namespace arm */ -using KwsClassifier = arm::app::Classifier; - enum opcodes { MENU_OPT_RUN_INF_NEXT = 1, /* Run on next vector. */ @@ -118,9 +116,9 @@ void main_loop() caseContext.Set("asrFrameStride", arm::app::asr::g_FrameStride); caseContext.Set("asrScoreThreshold", arm::app::asr::g_ScoreThreshold); /* Normalised score threshold. */ - KwsClassifier kwsClassifier; /* Classifier wrapper object. */ + arm::app::KwsClassifier kwsClassifier; /* Classifier wrapper object. */ arm::app::AsrClassifier asrClassifier; /* Classifier wrapper object. */ - caseContext.Set("kwsClassifier", kwsClassifier); + caseContext.Set("kwsClassifier", kwsClassifier); caseContext.Set("asrClassifier", asrClassifier); std::vector asrLabels; diff --git a/source/use_case/kws_asr/src/UseCaseHandler.cc b/source/use_case/kws_asr/src/UseCaseHandler.cc index 9427ae0..c5e6ad3 100644 --- a/source/use_case/kws_asr/src/UseCaseHandler.cc +++ b/source/use_case/kws_asr/src/UseCaseHandler.cc @@ -143,11 +143,8 @@ namespace app { while (audioDataSlider.HasNext()) { const int16_t* inferenceWindow = audioDataSlider.Next(); - /* The first window does not have cache ready. */ - preProcess.m_audioWindowIndex = audioDataSlider.Index(); - /* Run the pre-processing, inference and post-processing. */ - if (!preProcess.DoPreProcess(inferenceWindow, audio::MicroNetKwsMFCC::ms_defaultSamplingFreq)) { + if (!preProcess.DoPreProcess(inferenceWindow, audioDataSlider.Index())) { printf_err("KWS Pre-processing failed."); return output; } -- cgit v1.2.1