summaryrefslogtreecommitdiff
path: root/source/use_case/ad
diff options
context:
space:
mode:
Diffstat (limited to 'source/use_case/ad')
-rw-r--r--source/use_case/ad/include/AdMelSpectrogram.hpp4
-rw-r--r--source/use_case/ad/include/AdModel.hpp4
-rw-r--r--source/use_case/ad/include/MelSpectrogram.hpp22
-rw-r--r--source/use_case/ad/src/AdMelSpectrogram.cc17
-rw-r--r--source/use_case/ad/src/MelSpectrogram.cc53
-rw-r--r--source/use_case/ad/src/UseCaseHandler.cc113
6 files changed, 112 insertions, 101 deletions
diff --git a/source/use_case/ad/include/AdMelSpectrogram.hpp b/source/use_case/ad/include/AdMelSpectrogram.hpp
index cf8a1d4..30a77c1 100644
--- a/source/use_case/ad/include/AdMelSpectrogram.hpp
+++ b/source/use_case/ad/include/AdMelSpectrogram.hpp
@@ -60,8 +60,8 @@ namespace audio {
virtual bool ApplyMelFilterBank(
std::vector<float>& fftVec,
std::vector<std::vector<float>>& melFilterBank,
- std::vector<int32_t>& filterBankFilterFirst,
- std::vector<int32_t>& filterBankFilterLast,
+ std::vector<uint32_t>& filterBankFilterFirst,
+ std::vector<uint32_t>& filterBankFilterLast,
std::vector<float>& melEnergies) override;
/**
diff --git a/source/use_case/ad/include/AdModel.hpp b/source/use_case/ad/include/AdModel.hpp
index 2d83455..bbdf91c 100644
--- a/source/use_case/ad/include/AdModel.hpp
+++ b/source/use_case/ad/include/AdModel.hpp
@@ -41,10 +41,10 @@ namespace app {
private:
/* Maximum number of individual operations that can be enlisted */
- static constexpr int _ms_maxOpCnt = 6;
+ static constexpr int ms_maxOpCnt = 6;
/* A mutable op resolver instance */
- tflite::MicroMutableOpResolver<_ms_maxOpCnt> _m_opResolver;
+ tflite::MicroMutableOpResolver<ms_maxOpCnt> _m_opResolver;
};
} /* namespace app */
diff --git a/source/use_case/ad/include/MelSpectrogram.hpp b/source/use_case/ad/include/MelSpectrogram.hpp
index c1dd61e..22b5d29 100644
--- a/source/use_case/ad/include/MelSpectrogram.hpp
+++ b/source/use_case/ad/include/MelSpectrogram.hpp
@@ -49,7 +49,7 @@ namespace audio {
~MelSpecParams() = default;
/** @brief String representation of parameters */
- std::string Str();
+ std::string Str() const;
};
/**
@@ -76,7 +76,7 @@ namespace audio {
* @brief Constructor
* @param[in] params - Mel Spectrogram parameters
*/
- MelSpectrogram(const MelSpecParams& params);
+ explicit MelSpectrogram(const MelSpecParams& params);
MelSpectrogram() = delete;
~MelSpectrogram() = default;
@@ -148,7 +148,7 @@ namespace audio {
* bank weights and adding them up to be placed into
* bins, according to the filter bank's first and last
* indices (pre-computed for each filter bank element
- * by _CreateMelFilterBank function).
+ * by CreateMelFilterBank function).
* @param[in] fftVec Vector populated with FFT magnitudes
* @param[in] melFilterBank 2D Vector with filter bank weights
* @param[in] filterBankFilterFirst Vector containing the first indices of filter bank
@@ -162,8 +162,8 @@ namespace audio {
virtual bool ApplyMelFilterBank(
std::vector<float>& fftVec,
std::vector<std::vector<float>>& melFilterBank,
- std::vector<int32_t>& filterBankFilterFirst,
- std::vector<int32_t>& filterBankFilterLast,
+ std::vector<uint32_t>& filterBankFilterFirst,
+ std::vector<uint32_t>& filterBankFilterLast,
std::vector<float>& melEnergies);
/**
@@ -195,33 +195,33 @@ namespace audio {
std::vector<float> _m_melEnergies;
std::vector<float> _m_windowFunc;
std::vector<std::vector<float>> _m_melFilterBank;
- std::vector<int32_t> _m_filterBankFilterFirst;
- std::vector<int32_t> _m_filterBankFilterLast;
+ std::vector<uint32_t> _m_filterBankFilterFirst;
+ std::vector<uint32_t> _m_filterBankFilterLast;
bool _m_filterBankInitialised;
arm::app::math::FftInstance _m_fftInstance;
/**
* @brief Initialises the filter banks.
**/
- void _InitMelFilterBank();
+ void InitMelFilterBank();
/**
* @brief Signals whether the instance of MelSpectrogram has had its
* required buffers initialised
* @return True if initialised, false otherwise
**/
- bool _IsMelFilterBankInited();
+ bool IsMelFilterBankInited() const;
/**
* @brief Create mel filter banks for Mel Spectrogram calculation.
* @return 2D vector of floats
**/
- std::vector<std::vector<float>> _CreateMelFilterBank();
+ std::vector<std::vector<float>> CreateMelFilterBank();
/**
* @brief Computes the magnitude from an interleaved complex array
**/
- void _ConvertToPowerSpectrum();
+ void ConvertToPowerSpectrum();
};
diff --git a/source/use_case/ad/src/AdMelSpectrogram.cc b/source/use_case/ad/src/AdMelSpectrogram.cc
index 183c05c..e070eb8 100644
--- a/source/use_case/ad/src/AdMelSpectrogram.cc
+++ b/source/use_case/ad/src/AdMelSpectrogram.cc
@@ -18,6 +18,8 @@
#include "PlatformMath.hpp"
+#include <cfloat>
+
namespace arm {
namespace app {
namespace audio {
@@ -25,8 +27,8 @@ namespace audio {
bool AdMelSpectrogram::ApplyMelFilterBank(
std::vector<float>& fftVec,
std::vector<std::vector<float>>& melFilterBank,
- std::vector<int32_t>& filterBankFilterFirst,
- std::vector<int32_t>& filterBankFilterLast,
+ std::vector<uint32_t>& filterBankFilterFirst,
+ std::vector<uint32_t>& filterBankFilterLast,
std::vector<float>& melEnergies)
{
const size_t numBanks = melEnergies.size();
@@ -39,11 +41,12 @@ namespace audio {
for (size_t bin = 0; bin < numBanks; ++bin) {
auto filterBankIter = melFilterBank[bin].begin();
- float melEnergy = 1e-10; /* Avoid log of zero at later stages. */
- const int32_t firstIndex = filterBankFilterFirst[bin];
- const int32_t lastIndex = filterBankFilterLast[bin];
+ auto end = melFilterBank[bin].end();
+ float melEnergy = FLT_MIN; /* Avoid log of zero at later stages. */
+ const uint32_t firstIndex = filterBankFilterFirst[bin];
+ const uint32_t lastIndex = std::min<int32_t>(filterBankFilterLast[bin], fftVec.size() - 1);
- for (int32_t i = firstIndex; i <= lastIndex; ++i) {
+ for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) {
melEnergy += (*filterBankIter++ * fftVec[i]);
}
@@ -69,7 +72,7 @@ namespace audio {
/* Scale the log values. */
for (auto iterM = melEnergies.begin(), iterL = vecLogEnergies.begin();
- iterM != melEnergies.end(); ++iterM, ++iterL) {
+ iterM != melEnergies.end() && iterL != vecLogEnergies.end(); ++iterM, ++iterL) {
*iterM = *iterL * multiplier;
}
diff --git a/source/use_case/ad/src/MelSpectrogram.cc b/source/use_case/ad/src/MelSpectrogram.cc
index 86d57e6..372ebd8 100644
--- a/source/use_case/ad/src/MelSpectrogram.cc
+++ b/source/use_case/ad/src/MelSpectrogram.cc
@@ -42,7 +42,7 @@ namespace audio {
m_useHtkMethod(useHtkMethod)
{}
- std::string MelSpecParams::Str()
+ std::string MelSpecParams::Str() const
{
char strC[1024];
snprintf(strC, sizeof(strC) - 1, "\n \
@@ -71,7 +71,7 @@ namespace audio {
this->_m_params.m_numFbankBins, 0.0);
this->_m_windowFunc = std::vector<float>(this->_m_params.m_frameLen);
- const float multiplier = 2 * M_PI / this->_m_params.m_frameLen;
+ const auto multiplier = static_cast<float>(2 * M_PI / this->_m_params.m_frameLen);
/* Create window function. */
for (size_t i = 0; i < this->_m_params.m_frameLen; ++i) {
@@ -85,7 +85,7 @@ namespace audio {
void MelSpectrogram::Init()
{
- this->_InitMelFilterBank();
+ this->InitMelFilterBank();
}
float MelSpectrogram::MelScale(const float freq, const bool useHTKMethod)
@@ -121,8 +121,8 @@ namespace audio {
bool MelSpectrogram::ApplyMelFilterBank(
std::vector<float>& fftVec,
std::vector<std::vector<float>>& melFilterBank,
- std::vector<int32_t>& filterBankFilterFirst,
- std::vector<int32_t>& filterBankFilterLast,
+ std::vector<uint32_t>& filterBankFilterFirst,
+ std::vector<uint32_t>& filterBankFilterLast,
std::vector<float>& melEnergies)
{
const size_t numBanks = melEnergies.size();
@@ -135,11 +135,12 @@ namespace audio {
for (size_t bin = 0; bin < numBanks; ++bin) {
auto filterBankIter = melFilterBank[bin].begin();
+ auto end = melFilterBank[bin].end();
float melEnergy = FLT_MIN; /* Avoid log of zero at later stages */
- int32_t firstIndex = filterBankFilterFirst[bin];
- int32_t lastIndex = filterBankFilterLast[bin];
+ const uint32_t firstIndex = filterBankFilterFirst[bin];
+ const uint32_t lastIndex = std::min<int32_t>(filterBankFilterLast[bin], fftVec.size() - 1);
- for (int i = firstIndex; i <= lastIndex; ++i) {
+ for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) {
float energyRep = math::MathUtils::SqrtF32(fftVec[i]);
melEnergy += (*filterBankIter++ * energyRep);
}
@@ -152,14 +153,14 @@ namespace audio {
void MelSpectrogram::ConvertToLogarithmicScale(std::vector<float>& melEnergies)
{
- for (size_t bin = 0; bin < melEnergies.size(); ++bin) {
- melEnergies[bin] = logf(melEnergies[bin]);
+ for (float& melEnergy : melEnergies) {
+ melEnergy = logf(melEnergy);
}
}
- void MelSpectrogram::_ConvertToPowerSpectrum()
+ void MelSpectrogram::ConvertToPowerSpectrum()
{
- const uint32_t halfDim = this->_m_params.m_frameLenPadded / 2;
+ const uint32_t halfDim = this->_m_buffer.size() / 2;
/* Handle this special case. */
float firstEnergy = this->_m_buffer[0] * this->_m_buffer[0];
@@ -188,22 +189,22 @@ namespace audio {
return 1.f;
}
- void MelSpectrogram::_InitMelFilterBank()
+ void MelSpectrogram::InitMelFilterBank()
{
- if (!this->_IsMelFilterBankInited()) {
- this->_m_melFilterBank = this->_CreateMelFilterBank();
+ if (!this->IsMelFilterBankInited()) {
+ this->_m_melFilterBank = this->CreateMelFilterBank();
this->_m_filterBankInitialised = true;
}
}
- bool MelSpectrogram::_IsMelFilterBankInited()
+ bool MelSpectrogram::IsMelFilterBankInited() const
{
return this->_m_filterBankInitialised;
}
std::vector<float> MelSpectrogram::ComputeMelSpec(const std::vector<int16_t>& audioData, float trainingMean)
{
- this->_InitMelFilterBank();
+ this->InitMelFilterBank();
/* TensorFlow way of normalizing .wav data to (-1, 1). */
constexpr float normaliser = 1.0/(1<<15);
@@ -223,7 +224,7 @@ namespace audio {
math::MathUtils::FftF32(this->_m_frame, this->_m_buffer, this->_m_fftInstance);
/* Convert to power spectrum. */
- this->_ConvertToPowerSpectrum();
+ this->ConvertToPowerSpectrum();
/* Apply mel filterbanks. */
if (!this->ApplyMelFilterBank(this->_m_buffer,
@@ -245,7 +246,7 @@ namespace audio {
return this->_m_melEnergies;
}
- std::vector<std::vector<float>> MelSpectrogram::_CreateMelFilterBank()
+ std::vector<std::vector<float>> MelSpectrogram::CreateMelFilterBank()
{
size_t numFftBins = this->_m_params.m_frameLenPadded / 2;
float fftBinWidth = static_cast<float>(this->_m_params.m_samplingFreq) / this->_m_params.m_frameLenPadded;
@@ -260,17 +261,18 @@ namespace audio {
std::vector<std::vector<float>> melFilterBank(
this->_m_params.m_numFbankBins);
this->_m_filterBankFilterFirst =
- std::vector<int32_t>(this->_m_params.m_numFbankBins);
+ std::vector<uint32_t>(this->_m_params.m_numFbankBins);
this->_m_filterBankFilterLast =
- std::vector<int32_t>(this->_m_params.m_numFbankBins);
+ std::vector<uint32_t>(this->_m_params.m_numFbankBins);
for (size_t bin = 0; bin < this->_m_params.m_numFbankBins; bin++) {
float leftMel = melLowFreq + bin * melFreqDelta;
float centerMel = melLowFreq + (bin + 1) * melFreqDelta;
float rightMel = melLowFreq + (bin + 2) * melFreqDelta;
- int32_t firstIndex = -1;
- int32_t lastIndex = -1;
+ uint32_t firstIndex = 0;
+ uint32_t lastIndex = 0;
+ bool firstIndexFound = false;
const float normaliser = this->GetMelFilterBankNormaliser(leftMel, rightMel, this->_m_params.m_useHtkMethod);
for (size_t i = 0; i < numFftBins; ++i) {
@@ -287,8 +289,9 @@ namespace audio {
}
thisBin[i] = weight * normaliser;
- if (firstIndex == -1) {
+ if (!firstIndexFound) {
firstIndex = i;
+ firstIndexFound = true;
}
lastIndex = i;
}
@@ -298,7 +301,7 @@ namespace audio {
this->_m_filterBankFilterLast[bin] = lastIndex;
/* Copy the part we care about. */
- for (int32_t i = firstIndex; i <= lastIndex; ++i) {
+ for (uint32_t i = firstIndex; i <= lastIndex; ++i) {
melFilterBank[bin].push_back(thisBin[i]);
}
}
diff --git a/source/use_case/ad/src/UseCaseHandler.cc b/source/use_case/ad/src/UseCaseHandler.cc
index 1c15595..8f86966 100644
--- a/source/use_case/ad/src/UseCaseHandler.cc
+++ b/source/use_case/ad/src/UseCaseHandler.cc
@@ -32,7 +32,7 @@ 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);
+ static void IncrementAppCtxClipIdx(ApplicationContext& ctx);
/**
* @brief Helper function to set the audio clip index
@@ -40,7 +40,7 @@ namespace app {
* @param[in] idx value to be set
* @return true if index is set, false otherwise
**/
- static bool _SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx);
+ static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx);
/**
* @brief Presents inference results using the data presentation
@@ -50,7 +50,7 @@ namespace app {
* @param[in] threhsold 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);
+ static bool PresentInferenceResult(hal_platform& platform, float result, float threshold);
/**
* @brief Returns a function to perform feature calculation and populates input tensor data with
@@ -87,7 +87,7 @@ namespace app {
/* If the request has a valid size, set the audio index */
if (clipIndex < NUMBER_OF_FILES) {
- if (!_SetAppCtxClipIdx(ctx, clipIndex)) {
+ if (!SetAppCtxClipIdx(ctx, clipIndex)) {
return false;
}
}
@@ -216,20 +216,20 @@ namespace app {
dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0);
ctx.Set<float>("result", result);
- if (!_PresentInferenceResult(platform, result, scoreThreshold)) {
+ if (!PresentInferenceResult(platform, result, scoreThreshold)) {
return false;
}
profiler.PrintProfilingResult();
- _IncrementAppCtxClipIdx(ctx);
+ IncrementAppCtxClipIdx(ctx);
} while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx);
return true;
}
- static void _IncrementAppCtxClipIdx(ApplicationContext& ctx)
+ static void IncrementAppCtxClipIdx(ApplicationContext& ctx)
{
auto curAudioIdx = ctx.Get<uint32_t>("clipIndex");
@@ -241,7 +241,7 @@ namespace app {
ctx.Set<uint32_t>("clipIndex", curAudioIdx);
}
- static bool _SetAppCtxClipIdx(ApplicationContext& ctx, const uint32_t idx)
+ static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx)
{
if (idx >= NUMBER_OF_FILES) {
printf_err("Invalid idx %u (expected less than %u)\n",
@@ -252,7 +252,7 @@ namespace app {
return true;
}
- static bool _PresentInferenceResult(hal_platform& platform, float result, float threshold)
+ static bool PresentInferenceResult(hal_platform& platform, float result, float threshold)
{
constexpr uint32_t dataPsnTxtStartX1 = 20;
constexpr uint32_t dataPsnTxtStartY1 = 30;
@@ -275,7 +275,7 @@ namespace app {
platform.data_psn->present_data_text(
resultStr.c_str(), resultStr.size(),
- dataPsnTxtStartX1, rowIdx1, 0);
+ dataPsnTxtStartX1, rowIdx1, false);
info("%s\n", resultStr.c_str());
@@ -297,8 +297,8 @@ namespace app {
*/
template<class T>
std::function<void (std::vector<int16_t>&, size_t, bool, size_t, size_t)>
- _FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize,
- std::function<std::vector<T> (std::vector<int16_t>& )> compute)
+ 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);
@@ -335,24 +335,24 @@ namespace app {
}
template std::function<void (std::vector<int16_t>&, size_t , bool, size_t, size_t)>
- _FeatureCalc<int8_t>(TfLiteTensor* inputTensor,
- size_t cacheSize,
- std::function<std::vector<int8_t> (std::vector<int16_t>&)> compute);
+ 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, size_t)>
- _FeatureCalc<uint8_t>(TfLiteTensor* inputTensor,
- size_t cacheSize,
- std::function<std::vector<uint8_t> (std::vector<int16_t>&)> compute);
+ 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, size_t)>
- _FeatureCalc<int16_t>(TfLiteTensor* inputTensor,
- size_t cacheSize,
- std::function<std::vector<int16_t> (std::vector<int16_t>&)> compute);
+ 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, size_t)>
- _FeatureCalc<float>(TfLiteTensor *inputTensor,
- size_t cacheSize,
- std::function<std::vector<float>(std::vector<int16_t>&)> compute);
+ 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, size_t)>
@@ -370,38 +370,41 @@ namespace app {
switch (inputTensor->type) {
case kTfLiteInt8: {
- melSpecFeatureCalc = _FeatureCalc<int8_t>(inputTensor,
- cacheSize,
- [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
- return melSpec.MelSpecComputeQuant<int8_t>(audioDataWindow,
- quantScale,
- quantOffset,
- trainingMean);
- }
+ melSpecFeatureCalc = FeatureCalc<int8_t>(inputTensor,
+ cacheSize,
+ [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
+ return melSpec.MelSpecComputeQuant<int8_t>(
+ audioDataWindow,
+ quantScale,
+ quantOffset,
+ trainingMean);
+ }
);
break;
}
case kTfLiteUInt8: {
- melSpecFeatureCalc = _FeatureCalc<uint8_t>(inputTensor,
- cacheSize,
- [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
- return melSpec.MelSpecComputeQuant<uint8_t>(audioDataWindow,
- quantScale,
- quantOffset,
- trainingMean);
- }
+ melSpecFeatureCalc = FeatureCalc<uint8_t>(inputTensor,
+ cacheSize,
+ [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
+ return melSpec.MelSpecComputeQuant<uint8_t>(
+ audioDataWindow,
+ quantScale,
+ quantOffset,
+ trainingMean);
+ }
);
break;
}
case kTfLiteInt16: {
- melSpecFeatureCalc = _FeatureCalc<int16_t>(inputTensor,
- cacheSize,
- [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
- return melSpec.MelSpecComputeQuant<int16_t>(audioDataWindow,
- quantScale,
- quantOffset,
- trainingMean);
- }
+ melSpecFeatureCalc = FeatureCalc<int16_t>(inputTensor,
+ cacheSize,
+ [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
+ return melSpec.MelSpecComputeQuant<int16_t>(
+ audioDataWindow,
+ quantScale,
+ quantOffset,
+ trainingMean);
+ }
);
break;
}
@@ -411,12 +414,14 @@ namespace app {
} else {
- melSpecFeatureCalc = melSpecFeatureCalc = _FeatureCalc<float>(inputTensor,
- cacheSize,
- [=, &melSpec](std::vector<int16_t>& audioDataWindow) {
- return melSpec.ComputeMelSpec(audioDataWindow,
- trainingMean);
- });
+ melSpecFeatureCalc = melSpecFeatureCalc = FeatureCalc<float>(inputTensor,
+ cacheSize,
+ [=, &melSpec](
+ std::vector<int16_t>& audioDataWindow) {
+ return melSpec.ComputeMelSpec(
+ audioDataWindow,
+ trainingMean);
+ });
}
return melSpecFeatureCalc;
}
generated by cgit v1.2.1 (git 2.23.0) at 2024-06-26 02:46:25 +0000