/* * Copyright (c) 2021 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 "UseCaseHandler.hpp" #include "AdModel.hpp" #include "InputFiles.hpp" #include "Classifier.hpp" #include "hal.h" #include "AdMelSpectrogram.hpp" #include "AudioUtils.hpp" #include "UseCaseCommonUtils.hpp" #include "AdPostProcessing.hpp" namespace arm { namespace app { /** * @brief Helper function to increment current audio clip index * @param[in,out] ctx pointer to the application context object **/ static void IncrementAppCtxClipIdx(ApplicationContext& ctx); /** * @brief Helper function to set the audio clip index * @param[in,out] ctx pointer to the application context object * @param[in] idx value to be set * @return true if index is set, false otherwise **/ static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx); /** * @brief Presents inference results using the data presentation * object. * @param[in] platform reference to the hal platform object * @param[in] result average sum of classification results * @param[in] threshold if larger than this value we have an anomaly * @return true if successful, false otherwise **/ static bool PresentInferenceResult(hal_platform& platform, float result, float threshold); /** * @brief Returns a function to perform feature calculation and populates input tensor data with * MelSpe 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] melSpec MFCC feature calculator. * @param[in,out] inputTensor Input tensor pointer to store calculated features. * @param[in] cacheSize Size of the feture vectors cache (number of feature vectors). * @param[in] trainingMean Training mean. * @return function function to be called providing audio sample and sliding window index. */ static std::function&, int, bool, size_t, size_t)> GetFeatureCalculator(audio::AdMelSpectrogram& melSpec, TfLiteTensor* inputTensor, size_t cacheSize, float trainingMean); /* Vibration classification handler */ bool ClassifyVibrationHandler(ApplicationContext& ctx, uint32_t clipIndex, bool runAll) { auto& platform = ctx.Get("platform"); auto& profiler = ctx.Get("profiler"); constexpr uint32_t dataPsnTxtInfStartX = 20; constexpr uint32_t dataPsnTxtInfStartY = 40; platform.data_psn->clear(COLOR_BLACK); auto& model = ctx.Get("model"); /* If the request has a valid size, set the audio index */ if (clipIndex < NUMBER_OF_FILES) { if (!SetAppCtxClipIdx(ctx, clipIndex)) { return false; } } if (!model.IsInited()) { printf_err("Model is not initialised! Terminating processing.\n"); return false; } const auto frameLength = ctx.Get("frameLength"); const auto frameStride = ctx.Get("frameStride"); const auto scoreThreshold = ctx.Get("scoreThreshold"); const auto trainingMean = ctx.Get("trainingMean"); auto startClipIdx = ctx.Get("clipIndex"); TfLiteTensor* outputTensor = model.GetOutputTensor(0); TfLiteTensor* inputTensor = model.GetInputTensor(0); if (!inputTensor->dims) { printf_err("Invalid input tensor dims\n"); return false; } TfLiteIntArray* inputShape = model.GetInputShape(0); const uint32_t kNumRows = inputShape->data[1]; const uint32_t kNumCols = inputShape->data[2]; audio::AdMelSpectrogram melSpec = audio::AdMelSpectrogram(frameLength); melSpec.Init(); /* Deduce the data length required for 1 inference from the network parameters. */ const uint8_t inputResizeScale = 2; const uint32_t audioDataWindowSize = (((inputResizeScale * kNumCols) - 1) * frameStride) + frameLength; /* We are choosing to move by 20 frames across the audio for each inference. */ const uint8_t nMelSpecVectorsInAudioStride = 20; auto audioDataStride = nMelSpecVectorsInAudioStride * frameStride; do { auto currentIndex = ctx.Get("clipIndex"); /* Get the output index to look at based on id in the filename. */ int8_t machineOutputIndex = OutputIndexFromFileName(get_filename(currentIndex)); if (machineOutputIndex == -1) { return false; } /* Creating a Mel Spectrogram sliding window for the data required for 1 inference. * "resizing" done here by multiplying stride by resize scale. */ auto audioMelSpecWindowSlider = audio::SlidingWindow( get_audio_array(currentIndex), audioDataWindowSize, frameLength, frameStride * inputResizeScale); /* Creating a sliding window through the whole audio clip. */ auto audioDataSlider = audio::SlidingWindow( get_audio_array(currentIndex), get_audio_array_size(currentIndex), audioDataWindowSize, audioDataStride); /* Calculate number of the feature vectors in the window overlap region taking into account resizing. * These feature vectors will be reused.*/ auto numberOfReusedFeatureVectors = kNumRows - (nMelSpecVectorsInAudioStride / inputResizeScale); /* Construct feature calculation function. */ auto melSpecFeatureCalc = GetFeatureCalculator(melSpec, inputTensor, numberOfReusedFeatureVectors, trainingMean); if (!melSpecFeatureCalc){ return false; } /* Result is an averaged sum over inferences. */ float result = 0; /* Display message on the LCD - inference running. */ std::string str_inf{"Running inference... "}; platform.data_psn->present_data_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. */ audioMelSpecWindowSlider.Reset(inferenceWindow); /* The first window does not have cache ready. */ bool useCache = audioDataSlider.Index() > 0 && numberOfReusedFeatureVectors > 0; /* Start calculating features inside one audio sliding window. */ while (audioMelSpecWindowSlider.HasNext()) { const int16_t *melSpecWindow = audioMelSpecWindowSlider.Next(); std::vector melSpecAudioData = std::vector(melSpecWindow, melSpecWindow + frameLength); /* Compute features for this window and write them to input tensor. */ melSpecFeatureCalc(melSpecAudioData, audioMelSpecWindowSlider.Index(), useCache, nMelSpecVectorsInAudioStride, inputResizeScale); } info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, audioDataSlider.TotalStrides() + 1); /* Run inference over this audio clip sliding window */ if (!RunInference(model, profiler)) { return false; } /* Use the negative softmax score of the corresponding index as the outlier score */ std::vector dequantOutput = Dequantize(outputTensor); Softmax(dequantOutput); result += -dequantOutput[machineOutputIndex]; #if VERIFY_TEST_OUTPUT arm::app::DumpTensor(outputTensor); #endif /* VERIFY_TEST_OUTPUT */ } /* while (audioDataSlider.HasNext()) */ /* Use average over whole clip as final score. */ result /= (audioDataSlider.TotalStrides() + 1); /* Erase. */ str_inf = std::string(str_inf.size(), ' '); platform.data_psn->present_data_text( str_inf.c_str(), str_inf.size(), dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); ctx.Set("result", result); if (!PresentInferenceResult(platform, result, scoreThreshold)) { return false; } profiler.PrintProfilingResult(); IncrementAppCtxClipIdx(ctx); } while (runAll && ctx.Get("clipIndex") != startClipIdx); return true; } static void IncrementAppCtxClipIdx(ApplicationContext& ctx) { auto curAudioIdx = ctx.Get("clipIndex"); if (curAudioIdx + 1 >= NUMBER_OF_FILES) { ctx.Set("clipIndex", 0); return; } ++curAudioIdx; ctx.Set("clipIndex", curAudioIdx); } static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx) { if (idx >= NUMBER_OF_FILES) { printf_err("Invalid idx %" PRIu32 " (expected less than %u)\n", idx, NUMBER_OF_FILES); return false; } ctx.Set("clipIndex", idx); return true; } static bool PresentInferenceResult(hal_platform& platform, float result, float threshold) { constexpr uint32_t dataPsnTxtStartX1 = 20; constexpr uint32_t dataPsnTxtStartY1 = 30; constexpr uint32_t dataPsnTxtYIncr = 16; /* Row index increment */ platform.data_psn->set_text_color(COLOR_GREEN); /* Display each result */ uint32_t rowIdx1 = dataPsnTxtStartY1 + 2 * dataPsnTxtYIncr; std::string anomalyScore = std::string{"Average anomaly score is: "} + std::to_string(result); std::string anomalyThreshold = std::string("Anomaly threshold is: ") + std::to_string(threshold); std::string anomalyResult; if (result > threshold) { anomalyResult += std::string("Anomaly detected!"); } else { anomalyResult += std::string("Everything fine, no anomaly detected!"); } platform.data_psn->present_data_text( anomalyScore.c_str(), anomalyScore.size(), dataPsnTxtStartX1, rowIdx1, false); info("%s\n", anomalyScore.c_str()); info("%s\n", anomalyThreshold.c_str()); info("%s\n", anomalyResult.c_str()); 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 inputTensor model input tensor pointer. * @param cacheSize number of feature vectors to cache. Defined by the sliding window overlap. * @param compute features calculator function. * @return lambda function to compute features. */ template std::function&, size_t, bool, size_t, size_t)> FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, std::function (std::vector& )> compute) { /* Feature cache to be captured by lambda function*/ static std::vector> featureCache = std::vector>(cacheSize); return [=](std::vector& audioDataWindow, size_t index, bool useCache, size_t featuresOverlapIndex, size_t resizeScale) { T *tensorData = tflite::GetTensorData(inputTensor); std::vector 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() / resizeScale; auto sizeBytes = sizeof(T); /* Input should be transposed and "resized" by skipping elements. */ for (size_t outIndex = 0; outIndex < size; outIndex++) { std::memcpy(tensorData + (outIndex*size) + index, &features[outIndex*resizeScale], sizeBytes); } /* Start renewing cache as soon iteration goes out of the windows overlap. */ if (index >= featuresOverlapIndex / resizeScale) { featureCache[index - featuresOverlapIndex / resizeScale] = std::move(features); } }; } template std::function&, size_t , bool, size_t, size_t)> FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, std::function (std::vector&)> compute); template std::function&, size_t , bool, size_t, size_t)> FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, std::function (std::vector&)> compute); template std::function&, size_t , bool, size_t, size_t)> FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, std::function (std::vector&)> compute); template std::function&, size_t, bool, size_t, size_t)> FeatureCalc(TfLiteTensor *inputTensor, size_t cacheSize, std::function(std::vector&)> compute); static std::function&, int, bool, size_t, size_t)> GetFeatureCalculator(audio::AdMelSpectrogram& melSpec, TfLiteTensor* inputTensor, size_t cacheSize, float trainingMean) { std::function&, size_t, bool, size_t, size_t)> melSpecFeatureCalc; 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: { melSpecFeatureCalc = FeatureCalc(inputTensor, cacheSize, [=, &melSpec](std::vector& audioDataWindow) { return melSpec.MelSpecComputeQuant( audioDataWindow, quantScale, quantOffset, trainingMean); } ); break; } case kTfLiteUInt8: { melSpecFeatureCalc = FeatureCalc(inputTensor, cacheSize, [=, &melSpec](std::vector& audioDataWindow) { return melSpec.MelSpecComputeQuant( audioDataWindow, quantScale, quantOffset, trainingMean); } ); break; } case kTfLiteInt16: { melSpecFeatureCalc = FeatureCalc(inputTensor, cacheSize, [=, &melSpec](std::vector& audioDataWindow) { return melSpec.MelSpecComputeQuant( audioDataWindow, quantScale, quantOffset, trainingMean); } ); break; } default: printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); } } else { melSpecFeatureCalc = melSpecFeatureCalc = FeatureCalc(inputTensor, cacheSize, [=, &melSpec]( std::vector& audioDataWindow) { return melSpec.ComputeMelSpec( audioDataWindow, trainingMean); }); } return melSpecFeatureCalc; } } /* namespace app */ } /* namespace arm */