diff options
-rw-r--r-- | source/use_case/img_class/include/ImgClassProcessing.hpp | 23 | ||||
-rw-r--r-- | source/use_case/img_class/src/ImgClassProcessing.cc | 8 | ||||
-rw-r--r-- | source/use_case/img_class/src/UseCaseHandler.cc | 17 | ||||
-rw-r--r-- | source/use_case/kws/include/KwsProcessing.hpp | 135 | ||||
-rw-r--r-- | source/use_case/kws/src/KwsProcessing.cc | 220 | ||||
-rw-r--r-- | source/use_case/kws/src/UseCaseHandler.cc | 342 | ||||
-rw-r--r-- | tests/use_case/kws/KWSHandlerTest.cc | 2 |
7 files changed, 468 insertions, 279 deletions
diff --git a/source/use_case/img_class/include/ImgClassProcessing.hpp b/source/use_case/img_class/include/ImgClassProcessing.hpp index 5a59b5f..59db4a5 100644 --- a/source/use_case/img_class/include/ImgClassProcessing.hpp +++ b/source/use_case/img_class/include/ImgClassProcessing.hpp @@ -32,8 +32,19 @@ namespace app { class ImgClassPreProcess : public BasePreProcess { public: + /** + * @brief Constructor + * @param[in] model Pointer to the the Image classification Model object. + **/ explicit ImgClassPreProcess(Model* model); + /** + * @brief Should perform pre-processing of 'raw' input image data and load it into + * TFLite Micro input tensors ready for inference + * @param[in] input Pointer to the data that pre-processing will work on. + * @param[in] inputSize Size of the input data. + * @return true if successful, false otherwise. + **/ bool DoPreProcess(const void* input, size_t inputSize) override; }; @@ -50,10 +61,22 @@ namespace app { std::vector<ClassificationResult>& m_results; public: + /** + * @brief Constructor + * @param[in] classifier Classifier object used to get top N results from classification. + * @param[in] model Pointer to the the Image classification Model object. + * @param[in] labels Vector of string labels to identify each output of the model. + * @param[in] results Vector of classification results to store decoded outputs. + **/ ImgClassPostProcess(Classifier& classifier, Model* model, const std::vector<std::string>& labels, std::vector<ClassificationResult>& results); + /** + * @brief Should perform post-processing of the result of inference then populate + * populate classification result data for any later use. + * @return true if successful, false otherwise. + **/ bool DoPostProcess() override; }; diff --git a/source/use_case/img_class/src/ImgClassProcessing.cc b/source/use_case/img_class/src/ImgClassProcessing.cc index e33e3c1..6ba88ad 100644 --- a/source/use_case/img_class/src/ImgClassProcessing.cc +++ b/source/use_case/img_class/src/ImgClassProcessing.cc @@ -23,6 +23,9 @@ namespace app { ImgClassPreProcess::ImgClassPreProcess(Model* model) { + if (!model->IsInited()) { + printf_err("Model is not initialised!.\n"); + } this->m_model = model; } @@ -35,7 +38,7 @@ namespace app { auto input = static_cast<const uint8_t*>(data); TfLiteTensor* inputTensor = this->m_model->GetInputTensor(0); - memcpy(inputTensor->data.data, input, inputSize); + std::memcpy(inputTensor->data.data, input, inputSize); debug("Input tensor populated \n"); if (this->m_model->IsDataSigned()) { @@ -52,6 +55,9 @@ namespace app { m_labels{labels}, m_results{results} { + if (!model->IsInited()) { + printf_err("Model is not initialised!.\n"); + } this->m_model = model; } diff --git a/source/use_case/img_class/src/UseCaseHandler.cc b/source/use_case/img_class/src/UseCaseHandler.cc index 98e2b59..11a1aa8 100644 --- a/source/use_case/img_class/src/UseCaseHandler.cc +++ b/source/use_case/img_class/src/UseCaseHandler.cc @@ -37,6 +37,12 @@ namespace app { { auto& profiler = ctx.Get<Profiler&>("profiler"); auto& model = ctx.Get<Model&>("model"); + /* If the request has a valid size, set the image index as it might not be set. */ + if (imgIndex < NUMBER_OF_FILES) { + if (!SetAppCtxIfmIdx(ctx, imgIndex, "imgIndex")) { + return false; + } + } auto initialImIdx = ctx.Get<uint32_t>("imgIndex"); constexpr uint32_t dataPsnImgDownscaleFactor = 2; @@ -46,12 +52,7 @@ namespace app { constexpr uint32_t dataPsnTxtInfStartX = 150; constexpr uint32_t dataPsnTxtInfStartY = 40; - /* If the request has a valid size, set the image index. */ - if (imgIndex < NUMBER_OF_FILES) { - if (!SetAppCtxIfmIdx(ctx, imgIndex, "imgIndex")) { - return false; - } - } + if (!model.IsInited()) { printf_err("Model is not initialised! Terminating processing.\n"); return false; @@ -102,7 +103,7 @@ namespace app { /* Display message on the LCD - inference running. */ hal_lcd_display_text(str_inf.c_str(), str_inf.size(), - dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); + dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); /* Select the image to run inference with. */ info("Running inference on image %" PRIu32 " => %s\n", ctx.Get<uint32_t>("imgIndex"), @@ -129,7 +130,7 @@ namespace app { /* Erase. */ str_inf = std::string(str_inf.size(), ' '); hal_lcd_display_text(str_inf.c_str(), str_inf.size(), - dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); + dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); /* Add results to context for access outside handler. */ ctx.Set<std::vector<ClassificationResult>>("results", results); diff --git a/source/use_case/kws/include/KwsProcessing.hpp b/source/use_case/kws/include/KwsProcessing.hpp new file mode 100644 index 0000000..abf20ab --- /dev/null +++ b/source/use_case/kws/include/KwsProcessing.hpp @@ -0,0 +1,135 @@ +/* + * 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_PROCESSING_HPP +#define KWS_PROCESSING_HPP + +#include <AudioUtils.hpp> +#include "BaseProcessing.hpp" +#include "Model.hpp" +#include "Classifier.hpp" +#include "MicroNetKwsMfcc.hpp" + +#include <functional> + +namespace arm { +namespace app { + + /** + * @brief Pre-processing class for Keyword Spotting use case. + * Implements methods declared by BasePreProcess and anything else needed + * to populate input tensors ready for inference. + */ + class KWSPreProcess : public BasePreProcess { + + public: + /** + * @brief Constructor + * @param[in] model Pointer to the the KWS Model object. + * @param[in] numFeatures How many MFCC features to use. + * @param[in] mfccFrameLength Number of audio samples used to calculate one set of MFCC values when + * sliding a window through the audio sample. + * @param[in] mfccFrameStride Number of audio samples between consecutive windows. + **/ + explicit KWSPreProcess(Model* model, size_t numFeatures, int mfccFrameLength, int mfccFrameStride); + + /** + * @brief Should perform pre-processing of 'raw' input audio data and load it into + * TFLite Micro input tensors ready for inference. + * @param[in] input Pointer to the data that pre-processing will work on. + * @param[in] inputSize Size of the input data. + * @return true if successful, false otherwise. + **/ + bool DoPreProcess(const void* input, size_t inputSize) 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. */ + + private: + const int m_mfccFrameLength; + const int m_mfccFrameStride; + + audio::MicroNetKwsMFCC m_mfcc; + audio::SlidingWindow<const int16_t> m_mfccSlidingWindow; + size_t m_numMfccVectorsInAudioStride; + size_t m_numReusedMfccVectors; + std::function<void (std::vector<int16_t>&, int, bool, size_t)> m_mfccFeatureCalculator; + + /** + * @brief Returns a function to perform feature calculation and populates input tensor data with + * MFCC data. + * + * Input tensor data type check is performed to choose correct MFCC feature data type. + * If tensor has an integer data type then original features are quantised. + * + * Warning: MFCC calculator provided as input must have the same life scope as returned function. + * + * @param[in] mfcc MFCC feature calculator. + * @param[in,out] inputTensor Input tensor pointer to store calculated features. + * @param[in] cacheSize Size of the feature vectors cache (number of feature vectors). + * @return Function to be called providing audio sample and sliding window index. + */ + std::function<void (std::vector<int16_t>&, int, bool, size_t)> + GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, + TfLiteTensor* inputTensor, + size_t cacheSize); + + template<class T> + std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> + FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, + std::function<std::vector<T> (std::vector<int16_t>& )> compute); + }; + + /** + * @brief Post-processing class for Keyword Spotting use case. + * Implements methods declared by BasePostProcess and anything else needed + * to populate result vector. + */ + class KWSPostProcess : public BasePostProcess { + + private: + Classifier& m_kwsClassifier; + const std::vector<std::string>& m_labels; + std::vector<ClassificationResult>& m_results; + + public: + const float m_scoreThreshold; + /** + * @brief Constructor + * @param[in] classifier Classifier object used to get top N results from classification. + * @param[in] model Pointer to the the Image classification Model object. + * @param[in] labels Vector of string labels to identify each output of the model. + * @param[in] results Vector of classification results to store decoded outputs. + * @param[in] scoreThreshold Predicted model score must be larger than this value to be accepted. + **/ + KWSPostProcess(Classifier& classifier, Model* model, + const std::vector<std::string>& labels, + std::vector<ClassificationResult>& results, + float scoreThreshold); + + /** + * @brief Should perform post-processing of the result of inference then populate + * populate KWS result data for any later use. + * @return true if successful, false otherwise. + **/ + bool DoPostProcess() override; + }; + +} /* namespace app */ +} /* namespace arm */ + +#endif /* KWS_PROCESSING_HPP */
\ No newline at end of file diff --git a/source/use_case/kws/src/KwsProcessing.cc b/source/use_case/kws/src/KwsProcessing.cc new file mode 100644 index 0000000..b6b230c --- /dev/null +++ b/source/use_case/kws/src/KwsProcessing.cc @@ -0,0 +1,220 @@ +/* + * 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 "KwsProcessing.hpp" +#include "ImageUtils.hpp" +#include "log_macros.h" +#include "MicroNetKwsModel.hpp" + +namespace arm { +namespace app { + + KWSPreProcess::KWSPreProcess(Model* model, size_t numFeatures, int mfccFrameLength, int mfccFrameStride): + m_mfccFrameLength{mfccFrameLength}, + m_mfccFrameStride{mfccFrameStride}, + m_mfcc{audio::MicroNetKwsMFCC(numFeatures, mfccFrameLength)} + { + if (!model->IsInited()) { + printf_err("Model is not initialised!.\n"); + } + this->m_model = model; + this->m_mfcc.Init(); + + TfLiteIntArray* inputShape = model->GetInputShape(0); + const uint32_t numMfccFrames = inputShape->data[arm::app::MicroNetKwsModel::ms_inputRowsIdx]; + + /* Deduce the data length required for 1 inference from the network parameters. */ + this->m_audioDataWindowSize = numMfccFrames * this->m_mfccFrameStride + + (this->m_mfccFrameLength - this->m_mfccFrameStride); + + /* Creating an MFCC feature sliding window for the data required for 1 inference. */ + this->m_mfccSlidingWindow = audio::SlidingWindow<const int16_t>(nullptr, this->m_audioDataWindowSize, + this->m_mfccFrameLength, this->m_mfccFrameStride); + + /* For longer audio clips we choose to move by half the audio window size + * => for a 1 second window size there is an overlap of 0.5 seconds. */ + this->m_audioDataStride = this->m_audioDataWindowSize / 2; + + /* To have the previously calculated features re-usable, stride must be multiple + * of MFCC features window stride. Reduce stride through audio if needed. */ + if (0 != this->m_audioDataStride % this->m_mfccFrameStride) { + this->m_audioDataStride -= this->m_audioDataStride % this->m_mfccFrameStride; + } + + this->m_numMfccVectorsInAudioStride = this->m_audioDataStride / this->m_mfccFrameStride; + + /* Calculate number of the feature vectors in the window overlap region. + * These feature vectors will be reused.*/ + this->m_numReusedMfccVectors = this->m_mfccSlidingWindow.TotalStrides() + 1 + - this->m_numMfccVectorsInAudioStride; + + /* Construct feature calculation function. */ + this->m_mfccFeatureCalculator = GetFeatureCalculator(this->m_mfcc, this->m_model->GetInputTensor(0), + this->m_numReusedMfccVectors); + + if (!this->m_mfccFeatureCalculator) { + printf_err("Feature calculator not initialized."); + } + } + + bool KWSPreProcess::DoPreProcess(const void* data, size_t inputSize) + { + UNUSED(inputSize); + if (data == nullptr) { + printf_err("Data pointer is null"); + } + + /* Set the features sliding window to the new address. */ + auto input = static_cast<const int16_t*>(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; + + /* Use a sliding window to calculate MFCC features frame by frame. */ + while (this->m_mfccSlidingWindow.HasNext()) { + const int16_t* mfccWindow = this->m_mfccSlidingWindow.Next(); + + std::vector<int16_t> mfccFrameAudioData = std::vector<int16_t>(mfccWindow, + mfccWindow + this->m_mfccFrameLength); + + /* Compute features for this window and write them to input tensor. */ + this->m_mfccFeatureCalculator(mfccFrameAudioData, this->m_mfccSlidingWindow.Index(), + useCache, this->m_numMfccVectorsInAudioStride); + } + + debug("Input tensor populated \n"); + + return true; + } + + /** + * @brief Generic feature calculator factory. + * + * Returns lambda function to compute features using features cache. + * Real features math is done by a lambda function provided as a parameter. + * Features are written to input tensor memory. + * + * @tparam T Feature vector type. + * @param[in] inputTensor Model input tensor pointer. + * @param[in] cacheSize Number of feature vectors to cache. Defined by the sliding window overlap. + * @param[in] compute Features calculator function. + * @return Lambda function to compute features. + */ + template<class T> + std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> + KWSPreProcess::FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, + std::function<std::vector<T> (std::vector<int16_t>& )> compute) + { + /* Feature cache to be captured by lambda function. */ + static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize); + + return [=](std::vector<int16_t>& audioDataWindow, + size_t index, + bool useCache, + size_t featuresOverlapIndex) + { + T* tensorData = tflite::GetTensorData<T>(inputTensor); + std::vector<T> features; + + /* Reuse features from cache if cache is ready and sliding windows overlap. + * Overlap is in the beginning of sliding window with a size of a feature cache. */ + if (useCache && index < featureCache.size()) { + features = std::move(featureCache[index]); + } else { + features = std::move(compute(audioDataWindow)); + } + auto size = features.size(); + auto sizeBytes = sizeof(T) * size; + std::memcpy(tensorData + (index * size), features.data(), sizeBytes); + + /* Start renewing cache as soon iteration goes out of the windows overlap. */ + if (index >= featuresOverlapIndex) { + featureCache[index - featuresOverlapIndex] = std::move(features); + } + }; + } + + template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> + KWSPreProcess::FeatureCalc<int8_t>(TfLiteTensor* inputTensor, + size_t cacheSize, + std::function<std::vector<int8_t> (std::vector<int16_t>&)> compute); + + template std::function<void(std::vector<int16_t>&, size_t, bool, size_t)> + KWSPreProcess::FeatureCalc<float>(TfLiteTensor* inputTensor, + size_t cacheSize, + std::function<std::vector<float>(std::vector<int16_t>&)> compute); + + + std::function<void (std::vector<int16_t>&, int, bool, size_t)> + KWSPreProcess::GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, TfLiteTensor* inputTensor, size_t cacheSize) + { + std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> mfccFeatureCalc; + + TfLiteQuantization quant = inputTensor->quantization; + + if (kTfLiteAffineQuantization == quant.type) { + auto *quantParams = (TfLiteAffineQuantization *) quant.params; + const float quantScale = quantParams->scale->data[0]; + const int quantOffset = quantParams->zero_point->data[0]; + + switch (inputTensor->type) { + case kTfLiteInt8: { + mfccFeatureCalc = this->FeatureCalc<int8_t>(inputTensor, + cacheSize, + [=, &mfcc](std::vector<int16_t>& audioDataWindow) { + return mfcc.MfccComputeQuant<int8_t>(audioDataWindow, + quantScale, + quantOffset); + } + ); + break; + } + default: + printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); + } + } else { + mfccFeatureCalc = this->FeatureCalc<float>(inputTensor, cacheSize, + [&mfcc](std::vector<int16_t>& audioDataWindow) { + return mfcc.MfccCompute(audioDataWindow); } + ); + } + return mfccFeatureCalc; + } + + KWSPostProcess::KWSPostProcess(Classifier& classifier, Model* model, + const std::vector<std::string>& labels, + std::vector<ClassificationResult>& results, float scoreThreshold) + :m_kwsClassifier{classifier}, + m_labels{labels}, + m_results{results}, + m_scoreThreshold{scoreThreshold} + { + if (!model->IsInited()) { + printf_err("Model is not initialised!.\n"); + } + this->m_model = model; + } + + bool KWSPostProcess::DoPostProcess() + { + return this->m_kwsClassifier.GetClassificationResults( + this->m_model->GetOutputTensor(0), this->m_results, + this->m_labels, 1, true); + } + +} /* namespace app */ +} /* namespace arm */
\ No newline at end of file diff --git a/source/use_case/kws/src/UseCaseHandler.cc b/source/use_case/kws/src/UseCaseHandler.cc index e04cefc..350d34b 100644 --- a/source/use_case/kws/src/UseCaseHandler.cc +++ b/source/use_case/kws/src/UseCaseHandler.cc @@ -20,15 +20,14 @@ #include "Classifier.hpp" #include "MicroNetKwsModel.hpp" #include "hal.h" -#include "MicroNetKwsMfcc.hpp" #include "AudioUtils.hpp" #include "ImageUtils.hpp" #include "UseCaseCommonUtils.hpp" #include "KwsResult.hpp" #include "log_macros.h" +#include "KwsProcessing.hpp" #include <vector> -#include <functional> using KwsClassifier = arm::app::Classifier; @@ -37,36 +36,27 @@ namespace app { /** - * @brief Presents inference results using the data presentation - * object. - * @param[in] results Vector of classification results to be displayed. + * @brief Presents KWS inference results. + * @param[in] results Vector of KWS classification results to be displayed. * @return true if successful, false otherwise. **/ static bool PresentInferenceResult(const std::vector<arm::app::kws::KwsResult>& results); - /** - * @brief Returns a function to perform feature calculation and populates input tensor data with - * MFCC data. - * - * Input tensor data type check is performed to choose correct MFCC feature data type. - * If tensor has an integer data type then original features are quantised. - * - * Warning: MFCC calculator provided as input must have the same life scope as returned function. - * - * @param[in] mfcc MFCC feature calculator. - * @param[in,out] inputTensor Input tensor pointer to store calculated features. - * @param[in] cacheSize Size of the feature vectors cache (number of feature vectors). - * @return Function to be called providing audio sample and sliding window index. - */ - static std::function<void (std::vector<int16_t>&, int, bool, size_t)> - GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, - TfLiteTensor* inputTensor, - size_t cacheSize); - - /* Audio inference handler. */ + /* KWS inference handler. */ bool ClassifyAudioHandler(ApplicationContext& ctx, uint32_t clipIndex, bool runAll) { auto& profiler = ctx.Get<Profiler&>("profiler"); + auto& model = ctx.Get<Model&>("model"); + const auto mfccFrameLength = ctx.Get<int>("frameLength"); + const auto mfccFrameStride = ctx.Get<int>("frameStride"); + const auto scoreThreshold = ctx.Get<float>("scoreThreshold"); + /* If the request has a valid size, set the audio index. */ + if (clipIndex < NUMBER_OF_FILES) { + if (!SetAppCtxIfmIdx(ctx, clipIndex,"clipIndex")) { + return false; + } + } + auto initialClipIdx = ctx.Get<uint32_t>("clipIndex"); constexpr uint32_t dataPsnTxtInfStartX = 20; constexpr uint32_t dataPsnTxtInfStartY = 40; @@ -74,27 +64,13 @@ namespace app { (arm::app::MicroNetKwsModel::ms_inputRowsIdx > arm::app::MicroNetKwsModel::ms_inputColsIdx)? arm::app::MicroNetKwsModel::ms_inputRowsIdx : arm::app::MicroNetKwsModel::ms_inputColsIdx); - auto& model = ctx.Get<Model&>("model"); - /* If the request has a valid size, set the audio index. */ - if (clipIndex < NUMBER_OF_FILES) { - if (!SetAppCtxIfmIdx(ctx, clipIndex,"clipIndex")) { - return false; - } - } if (!model.IsInited()) { printf_err("Model is not initialised! Terminating processing.\n"); return false; } - const auto frameLength = ctx.Get<int>("frameLength"); - const auto frameStride = ctx.Get<int>("frameStride"); - const auto scoreThreshold = ctx.Get<float>("scoreThreshold"); - auto startClipIdx = ctx.Get<uint32_t>("clipIndex"); - - TfLiteTensor* outputTensor = model.GetOutputTensor(0); TfLiteTensor* inputTensor = model.GetInputTensor(0); - if (!inputTensor->dims) { printf_err("Invalid input tensor dims\n"); return false; @@ -103,130 +79,89 @@ namespace app { return false; } + /* Get input shape for feature extraction. */ TfLiteIntArray* inputShape = model.GetInputShape(0); - const uint32_t kNumCols = inputShape->data[arm::app::MicroNetKwsModel::ms_inputColsIdx]; - const uint32_t kNumRows = inputShape->data[arm::app::MicroNetKwsModel::ms_inputRowsIdx]; - - audio::MicroNetKwsMFCC mfcc = audio::MicroNetKwsMFCC(kNumCols, frameLength); - mfcc.Init(); + const uint32_t numMfccFeatures = inputShape->data[arm::app::MicroNetKwsModel::ms_inputColsIdx]; - /* Deduce the data length required for 1 inference from the network parameters. */ - auto audioDataWindowSize = kNumRows * frameStride + (frameLength - frameStride); - auto mfccWindowSize = frameLength; - auto mfccWindowStride = frameStride; - - /* We choose to move by half the window size => for a 1 second window size - * there is an overlap of 0.5 seconds. */ - auto audioDataStride = audioDataWindowSize / 2; - - /* To have the previously calculated features re-usable, stride must be multiple - * of MFCC features window stride. */ - if (0 != audioDataStride % mfccWindowStride) { + /* We expect to be sampling 1 second worth of data at a time. + * NOTE: This is only used for time stamp calculation. */ + const float secondsPerSample = 1.0 / audio::MicroNetKwsMFCC::ms_defaultSamplingFreq; - /* Reduce the stride. */ - audioDataStride -= audioDataStride % mfccWindowStride; - } + /* Set up pre and post-processing. */ + KWSPreProcess preprocess = KWSPreProcess(&model, numMfccFeatures, mfccFrameLength, mfccFrameStride); - auto nMfccVectorsInAudioStride = audioDataStride/mfccWindowStride; + std::vector<ClassificationResult> singleInfResult; + KWSPostProcess postprocess = KWSPostProcess(ctx.Get<KwsClassifier &>("classifier"), &model, + ctx.Get<std::vector<std::string>&>("labels"), + singleInfResult, scoreThreshold); - /* We expect to be sampling 1 second worth of data at a time. - * NOTE: This is only used for time stamp calculation. */ - const float secondsPerSample = 1.0/audio::MicroNetKwsMFCC::ms_defaultSamplingFreq; + UseCaseRunner runner = UseCaseRunner(&preprocess, &postprocess, &model); do { hal_lcd_clear(COLOR_BLACK); auto currentIndex = ctx.Get<uint32_t>("clipIndex"); - /* Creating a mfcc features sliding window for the data required for 1 inference. */ - auto audioMFCCWindowSlider = audio::SlidingWindow<const int16_t>( - get_audio_array(currentIndex), - audioDataWindowSize, mfccWindowSize, - mfccWindowStride); - /* Creating a sliding window through the whole audio clip. */ auto audioDataSlider = audio::SlidingWindow<const int16_t>( - get_audio_array(currentIndex), - get_audio_array_size(currentIndex), - audioDataWindowSize, audioDataStride); + get_audio_array(currentIndex), + get_audio_array_size(currentIndex), + preprocess.m_audioDataWindowSize, preprocess.m_audioDataStride); - /* Calculate number of the feature vectors in the window overlap region. - * These feature vectors will be reused.*/ - auto numberOfReusedFeatureVectors = audioMFCCWindowSlider.TotalStrides() + 1 - - nMfccVectorsInAudioStride; - - /* Construct feature calculation function. */ - auto mfccFeatureCalc = GetFeatureCalculator(mfcc, inputTensor, - numberOfReusedFeatureVectors); - - if (!mfccFeatureCalc){ - return false; - } - - /* Declare a container for results. */ - std::vector<arm::app::kws::KwsResult> results; + /* Declare a container to hold results from across the whole audio clip. */ + std::vector<kws::KwsResult> finalResults; /* Display message on the LCD - inference running. */ std::string str_inf{"Running inference... "}; - hal_lcd_display_text( - str_inf.c_str(), str_inf.size(), - dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); + hal_lcd_display_text(str_inf.c_str(), str_inf.size(), + dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); info("Running inference on audio clip %" PRIu32 " => %s\n", currentIndex, get_filename(currentIndex)); /* Start sliding through audio clip. */ while (audioDataSlider.HasNext()) { - const int16_t *inferenceWindow = audioDataSlider.Next(); - - /* We moved to the next window - set the features sliding to the new address. */ - audioMFCCWindowSlider.Reset(inferenceWindow); + const int16_t* inferenceWindow = audioDataSlider.Next(); /* The first window does not have cache ready. */ - bool useCache = audioDataSlider.Index() > 0 && numberOfReusedFeatureVectors > 0; - - /* Start calculating features inside one audio sliding window. */ - while (audioMFCCWindowSlider.HasNext()) { - const int16_t *mfccWindow = audioMFCCWindowSlider.Next(); - std::vector<int16_t> mfccAudioData = std::vector<int16_t>(mfccWindow, - mfccWindow + mfccWindowSize); - /* Compute features for this window and write them to input tensor. */ - mfccFeatureCalc(mfccAudioData, - audioMFCCWindowSlider.Index(), - useCache, - nMfccVectorsInAudioStride); - } + preprocess.m_audioWindowIndex = audioDataSlider.Index(); info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, audioDataSlider.TotalStrides() + 1); - /* Run inference over this audio clip sliding window. */ - if (!RunInference(model, profiler)) { + /* Run the pre-processing, inference and post-processing. */ + if (!runner.PreProcess(inferenceWindow, audio::MicroNetKwsMFCC::ms_defaultSamplingFreq)) { + return false; + } + + profiler.StartProfiling("Inference"); + if (!runner.RunInference()) { return false; } + profiler.StopProfiling(); - std::vector<ClassificationResult> classificationResult; - auto& classifier = ctx.Get<KwsClassifier&>("classifier"); - classifier.GetClassificationResults(outputTensor, classificationResult, - ctx.Get<std::vector<std::string>&>("labels"), 1, true); + if (!runner.PostProcess()) { + return false; + } - results.emplace_back(kws::KwsResult(classificationResult, - audioDataSlider.Index() * secondsPerSample * audioDataStride, - audioDataSlider.Index(), scoreThreshold)); + /* Add results from this window to our final results vector. */ + finalResults.emplace_back(kws::KwsResult(singleInfResult, + audioDataSlider.Index() * secondsPerSample * preprocess.m_audioDataStride, + audioDataSlider.Index(), postprocess.m_scoreThreshold)); #if VERIFY_TEST_OUTPUT + TfLiteTensor* outputTensor = model.GetOutputTensor(0); arm::app::DumpTensor(outputTensor); #endif /* VERIFY_TEST_OUTPUT */ } /* while (audioDataSlider.HasNext()) */ /* Erase. */ str_inf = std::string(str_inf.size(), ' '); - hal_lcd_display_text( - str_inf.c_str(), str_inf.size(), - dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); + hal_lcd_display_text(str_inf.c_str(), str_inf.size(), + dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); - ctx.Set<std::vector<arm::app::kws::KwsResult>>("results", results); + ctx.Set<std::vector<kws::KwsResult>>("results", finalResults); - if (!PresentInferenceResult(results)) { + if (!PresentInferenceResult(finalResults)) { return false; } @@ -234,58 +169,11 @@ namespace app { IncrementAppCtxIfmIdx(ctx,"clipIndex"); - } while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx); + } while (runAll && ctx.Get<uint32_t>("clipIndex") != initialClipIdx); return true; } - /** - * @brief Generic feature calculator factory. - * - * Returns lambda function to compute features using features cache. - * Real features math is done by a lambda function provided as a parameter. - * Features are written to input tensor memory. - * - * @tparam T Feature vector type. - * @param[in] inputTensor Model input tensor pointer. - * @param[in] cacheSize Number of feature vectors to cache. Defined by the sliding window overlap. - * @param[in] compute Features calculator function. - * @return Lambda function to compute features. - */ - template<class T> - std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> - FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, - std::function<std::vector<T> (std::vector<int16_t>& )> compute) - { - /* Feature cache to be captured by lambda function. */ - static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize); - - return [=](std::vector<int16_t>& audioDataWindow, - size_t index, - bool useCache, - size_t featuresOverlapIndex) - { - T *tensorData = tflite::GetTensorData<T>(inputTensor); - std::vector<T> features; - - /* Reuse features from cache if cache is ready and sliding windows overlap. - * Overlap is in the beginning of sliding window with a size of a feature cache. */ - if (useCache && index < featureCache.size()) { - features = std::move(featureCache[index]); - } else { - features = std::move(compute(audioDataWindow)); - } - auto size = features.size(); - auto sizeBytes = sizeof(T) * size; - std::memcpy(tensorData + (index * size), features.data(), sizeBytes); - - /* Start renewing cache as soon iteration goes out of the windows overlap. */ - if (index >= featuresOverlapIndex) { - featureCache[index - featuresOverlapIndex] = std::move(features); - } - }; - } - static bool PresentInferenceResult(const std::vector<arm::app::kws::KwsResult>& results) { constexpr uint32_t dataPsnTxtStartX1 = 20; @@ -299,40 +187,39 @@ namespace app { /* Display each result */ uint32_t rowIdx1 = dataPsnTxtStartY1 + 2 * dataPsnTxtYIncr; - for (uint32_t i = 0; i < results.size(); ++i) { + for (const auto & result : results) { std::string topKeyword{"<none>"}; float score = 0.f; - if (!results[i].m_resultVec.empty()) { - topKeyword = results[i].m_resultVec[0].m_label; - score = results[i].m_resultVec[0].m_normalisedVal; + if (!result.m_resultVec.empty()) { + topKeyword = result.m_resultVec[0].m_label; + score = result.m_resultVec[0].m_normalisedVal; } std::string resultStr = - std::string{"@"} + std::to_string(results[i].m_timeStamp) + + std::string{"@"} + std::to_string(result.m_timeStamp) + std::string{"s: "} + topKeyword + std::string{" ("} + std::to_string(static_cast<int>(score * 100)) + std::string{"%)"}; - hal_lcd_display_text( - resultStr.c_str(), resultStr.size(), + hal_lcd_display_text(resultStr.c_str(), resultStr.size(), dataPsnTxtStartX1, rowIdx1, false); rowIdx1 += dataPsnTxtYIncr; - if (results[i].m_resultVec.empty()) { + if (result.m_resultVec.empty()) { info("For timestamp: %f (inference #: %" PRIu32 "); label: %s; threshold: %f\n", - results[i].m_timeStamp, results[i].m_inferenceNumber, + result.m_timeStamp, result.m_inferenceNumber, topKeyword.c_str(), - results[i].m_threshold); + result.m_threshold); } else { - for (uint32_t j = 0; j < results[i].m_resultVec.size(); ++j) { + for (uint32_t j = 0; j < result.m_resultVec.size(); ++j) { info("For timestamp: %f (inference #: %" PRIu32 "); label: %s, score: %f; threshold: %f\n", - results[i].m_timeStamp, - results[i].m_inferenceNumber, - results[i].m_resultVec[j].m_label.c_str(), - results[i].m_resultVec[j].m_normalisedVal, - results[i].m_threshold); + result.m_timeStamp, + result.m_inferenceNumber, + result.m_resultVec[j].m_label.c_str(), + result.m_resultVec[j].m_normalisedVal, + result.m_threshold); } } } @@ -340,88 +227,5 @@ namespace app { return true; } - template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> - FeatureCalc<int8_t>(TfLiteTensor* inputTensor, - size_t cacheSize, - std::function<std::vector<int8_t> (std::vector<int16_t>& )> compute); - - template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> - FeatureCalc<uint8_t>(TfLiteTensor* inputTensor, - size_t cacheSize, - std::function<std::vector<uint8_t> (std::vector<int16_t>& )> compute); - - template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> - FeatureCalc<int16_t>(TfLiteTensor* inputTensor, - size_t cacheSize, - std::function<std::vector<int16_t> (std::vector<int16_t>& )> compute); - - template std::function<void(std::vector<int16_t>&, size_t, bool, size_t)> - FeatureCalc<float>(TfLiteTensor* inputTensor, - size_t cacheSize, - std::function<std::vector<float>(std::vector<int16_t>&)> compute); - - - static std::function<void (std::vector<int16_t>&, int, bool, size_t)> - GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, TfLiteTensor* inputTensor, size_t cacheSize) - { - std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> mfccFeatureCalc; - - TfLiteQuantization quant = inputTensor->quantization; - - if (kTfLiteAffineQuantization == quant.type) { - - auto *quantParams = (TfLiteAffineQuantization *) quant.params; - const float quantScale = quantParams->scale->data[0]; - const int quantOffset = quantParams->zero_point->data[0]; - - switch (inputTensor->type) { - case kTfLiteInt8: { - mfccFeatureCalc = FeatureCalc<int8_t>(inputTensor, - cacheSize, - [=, &mfcc](std::vector<int16_t>& audioDataWindow) { - return mfcc.MfccComputeQuant<int8_t>(audioDataWindow, - quantScale, - quantOffset); - } - ); - break; - } - case kTfLiteUInt8: { - mfccFeatureCalc = FeatureCalc<uint8_t>(inputTensor, - cacheSize, - [=, &mfcc](std::vector<int16_t>& audioDataWindow) { - return mfcc.MfccComputeQuant<uint8_t>(audioDataWindow, - quantScale, - quantOffset); - } - ); - break; - } - case kTfLiteInt16: { - mfccFeatureCalc = FeatureCalc<int16_t>(inputTensor, - cacheSize, - [=, &mfcc](std::vector<int16_t>& audioDataWindow) { - return mfcc.MfccComputeQuant<int16_t>(audioDataWindow, - quantScale, - quantOffset); - } - ); - break; - } - default: - printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); - } - - - } else { - mfccFeatureCalc = mfccFeatureCalc = FeatureCalc<float>(inputTensor, - cacheSize, - [&mfcc](std::vector<int16_t>& audioDataWindow) { - return mfcc.MfccCompute(audioDataWindow); - }); - } - return mfccFeatureCalc; - } - } /* namespace app */ -} /* namespace arm */
\ No newline at end of file +} /* namespace arm */ diff --git a/tests/use_case/kws/KWSHandlerTest.cc b/tests/use_case/kws/KWSHandlerTest.cc index d0a8a3f..c24faa4 100644 --- a/tests/use_case/kws/KWSHandlerTest.cc +++ b/tests/use_case/kws/KWSHandlerTest.cc @@ -67,7 +67,7 @@ TEST_CASE("Inference by index") auto checker = [&](uint32_t audioIndex, std::vector<uint32_t> labelIndex) { - caseContext.Set<uint32_t>("audioIndex", audioIndex); + caseContext.Set<uint32_t>("clipIndex", audioIndex); std::vector<std::string> labels; GetLabelsVector(labels); |