Richard Burton | 4e00279 | 2022-05-04 09:45:02 +0100 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2022 Arm Limited. All rights reserved. |
| 3 | * SPDX-License-Identifier: Apache-2.0 |
| 4 | * |
| 5 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | * you may not use this file except in compliance with the License. |
| 7 | * You may obtain a copy of the License at |
| 8 | * |
| 9 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | * |
| 11 | * Unless required by applicable law or agreed to in writing, software |
| 12 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | * See the License for the specific language governing permissions and |
| 15 | * limitations under the License. |
| 16 | */ |
| 17 | #ifndef AD_PROCESSING_HPP |
| 18 | #define AD_PROCESSING_HPP |
| 19 | |
| 20 | #include "BaseProcessing.hpp" |
| 21 | #include "AudioUtils.hpp" |
| 22 | #include "AdMelSpectrogram.hpp" |
| 23 | #include "log_macros.h" |
| 24 | |
| 25 | namespace arm { |
| 26 | namespace app { |
| 27 | |
| 28 | /** |
| 29 | * @brief Pre-processing class for anomaly detection use case. |
| 30 | * Implements methods declared by BasePreProcess and anything else needed |
| 31 | * to populate input tensors ready for inference. |
| 32 | */ |
| 33 | class AdPreProcess : public BasePreProcess { |
| 34 | |
| 35 | public: |
| 36 | /** |
| 37 | * @brief Constructor for AdPreProcess class objects |
| 38 | * @param[in] inputTensor input tensor pointer from the tensor arena. |
| 39 | * @param[in] melSpectrogramFrameLen MEL spectrogram's frame length |
| 40 | * @param[in] melSpectrogramFrameStride MEL spectrogram's frame stride |
| 41 | * @param[in] adModelTrainingMean Training mean for the Anomaly detection model being used. |
| 42 | */ |
| 43 | explicit AdPreProcess(TfLiteTensor* inputTensor, |
| 44 | uint32_t melSpectrogramFrameLen, |
| 45 | uint32_t melSpectrogramFrameStride, |
| 46 | float adModelTrainingMean); |
| 47 | |
| 48 | ~AdPreProcess() = default; |
| 49 | |
| 50 | /** |
| 51 | * @brief Function to invoke pre-processing and populate the input vector |
| 52 | * @param input pointer to input data. For anomaly detection, this is the pointer to |
| 53 | * the audio data. |
| 54 | * @param inputSize Size of the data being passed in for pre-processing. |
| 55 | * @return True if successful, false otherwise. |
| 56 | */ |
| 57 | bool DoPreProcess(const void* input, size_t inputSize) override; |
| 58 | |
| 59 | /** |
| 60 | * @brief Getter function for audio window size computed when constructing |
| 61 | * the class object. |
| 62 | * @return Audio window size as 32 bit unsigned integer. |
| 63 | */ |
| 64 | uint32_t GetAudioWindowSize(); |
| 65 | |
| 66 | /** |
| 67 | * @brief Getter function for audio window stride computed when constructing |
| 68 | * the class object. |
| 69 | * @return Audio window stride as 32 bit unsigned integer. |
| 70 | */ |
| 71 | uint32_t GetAudioDataStride(); |
| 72 | |
| 73 | /** |
| 74 | * @brief Setter function for current audio index. This is only used for evaluating |
| 75 | * if previously computed features can be re-used from cache. |
| 76 | */ |
| 77 | void SetAudioWindowIndex(uint32_t idx); |
| 78 | |
| 79 | private: |
| 80 | bool m_validInstance{false}; /**< Indicates the current object is valid. */ |
| 81 | uint32_t m_melSpectrogramFrameLen{}; /**< MEL spectrogram's window frame length */ |
| 82 | uint32_t m_melSpectrogramFrameStride{}; /**< MEL spectrogram's window frame stride */ |
| 83 | uint8_t m_inputResizeScale{}; /**< Downscaling factor for the MEL energy matrix. */ |
| 84 | uint32_t m_numMelSpecVectorsInAudioStride{}; /**< Number of frames to move across the audio. */ |
| 85 | uint32_t m_audioDataWindowSize{}; /**< Audio window size computed based on other parameters. */ |
| 86 | uint32_t m_audioDataStride{}; /**< Audio window stride computed. */ |
| 87 | uint32_t m_numReusedFeatureVectors{}; /**< Number of MEL vectors that can be re-used */ |
| 88 | uint32_t m_audioWindowIndex{}; /**< Current audio window index (from audio's sliding window) */ |
| 89 | |
| 90 | audio::SlidingWindow<const int16_t> m_melWindowSlider; /**< Internal MEL spectrogram window slider */ |
| 91 | audio::AdMelSpectrogram m_melSpec; /**< MEL spectrogram computation object */ |
| 92 | std::function<void |
| 93 | (std::vector<int16_t>&, int, bool, size_t, size_t)> m_featureCalc; /**< Feature calculator object */ |
| 94 | }; |
| 95 | |
| 96 | class AdPostProcess : public BasePostProcess { |
| 97 | public: |
| 98 | /** |
| 99 | * @brief Constructor for AdPostProcess object. |
| 100 | * @param[in] outputTensor Output tensor pointer. |
| 101 | */ |
| 102 | explicit AdPostProcess(TfLiteTensor* outputTensor); |
| 103 | |
| 104 | ~AdPostProcess() = default; |
| 105 | |
| 106 | /** |
| 107 | * @brief Function to do the post-processing on the output tensor. |
| 108 | * @return True if successful, false otherwise. |
| 109 | */ |
| 110 | bool DoPostProcess() override; |
| 111 | |
| 112 | /** |
| 113 | * @brief Getter function for an element from the de-quantised output vector. |
| 114 | * @param index Index of the element to be retrieved. |
| 115 | * @return index represented as a 32 bit floating point number. |
| 116 | */ |
| 117 | float GetOutputValue(uint32_t index); |
| 118 | |
| 119 | private: |
| 120 | TfLiteTensor* m_outputTensor{}; /**< Output tensor pointer */ |
| 121 | std::vector<float> m_dequantizedOutputVec{}; /**< Internal output vector */ |
| 122 | |
| 123 | /** |
| 124 | * @brief De-quantizes and flattens the output tensor into a vector. |
| 125 | * @tparam T template parameter to indicate data type. |
| 126 | * @return True if successful, false otherwise. |
| 127 | */ |
| 128 | template<typename T> |
| 129 | bool Dequantize() |
| 130 | { |
| 131 | TfLiteTensor* tensor = this->m_outputTensor; |
| 132 | if (tensor == nullptr) { |
| 133 | printf_err("Invalid output tensor.\n"); |
| 134 | return false; |
| 135 | } |
| 136 | T* tensorData = tflite::GetTensorData<T>(tensor); |
| 137 | |
| 138 | uint32_t totalOutputSize = 1; |
| 139 | for (int inputDim = 0; inputDim < tensor->dims->size; inputDim++){ |
| 140 | totalOutputSize *= tensor->dims->data[inputDim]; |
| 141 | } |
| 142 | |
| 143 | /* For getting the floating point values, we need quantization parameters */ |
| 144 | QuantParams quantParams = GetTensorQuantParams(tensor); |
| 145 | |
| 146 | this->m_dequantizedOutputVec = std::vector<float>(totalOutputSize, 0); |
| 147 | |
| 148 | for (size_t i = 0; i < totalOutputSize; ++i) { |
| 149 | this->m_dequantizedOutputVec[i] = quantParams.scale * (tensorData[i] - quantParams.offset); |
| 150 | } |
| 151 | |
| 152 | return true; |
| 153 | } |
| 154 | }; |
| 155 | |
| 156 | /* Templated instances available: */ |
| 157 | template bool AdPostProcess::Dequantize<int8_t>(); |
| 158 | |
| 159 | /** |
| 160 | * @brief Generic feature calculator factory. |
| 161 | * |
| 162 | * Returns lambda function to compute features using features cache. |
| 163 | * Real features math is done by a lambda function provided as a parameter. |
| 164 | * Features are written to input tensor memory. |
| 165 | * |
| 166 | * @tparam T feature vector type. |
| 167 | * @param inputTensor model input tensor pointer. |
| 168 | * @param cacheSize number of feature vectors to cache. Defined by the sliding window overlap. |
| 169 | * @param compute features calculator function. |
| 170 | * @return lambda function to compute features. |
| 171 | */ |
| 172 | template<class T> |
| 173 | std::function<void (std::vector<int16_t>&, size_t, bool, size_t, size_t)> |
| 174 | FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, |
| 175 | std::function<std::vector<T> (std::vector<int16_t>& )> compute) |
| 176 | { |
| 177 | /* Feature cache to be captured by lambda function*/ |
| 178 | static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize); |
| 179 | |
| 180 | return [=](std::vector<int16_t>& audioDataWindow, |
| 181 | size_t index, |
| 182 | bool useCache, |
| 183 | size_t featuresOverlapIndex, |
| 184 | size_t resizeScale) |
| 185 | { |
| 186 | T* tensorData = tflite::GetTensorData<T>(inputTensor); |
| 187 | std::vector<T> features; |
| 188 | |
| 189 | /* Reuse features from cache if cache is ready and sliding windows overlap. |
| 190 | * Overlap is in the beginning of sliding window with a size of a feature cache. */ |
| 191 | if (useCache && index < featureCache.size()) { |
| 192 | features = std::move(featureCache[index]); |
| 193 | } else { |
| 194 | features = std::move(compute(audioDataWindow)); |
| 195 | } |
| 196 | auto size = features.size() / resizeScale; |
| 197 | auto sizeBytes = sizeof(T); |
| 198 | |
| 199 | /* Input should be transposed and "resized" by skipping elements. */ |
| 200 | for (size_t outIndex = 0; outIndex < size; outIndex++) { |
| 201 | std::memcpy(tensorData + (outIndex*size) + index, &features[outIndex*resizeScale], sizeBytes); |
| 202 | } |
| 203 | |
| 204 | /* Start renewing cache as soon iteration goes out of the windows overlap. */ |
| 205 | if (index >= featuresOverlapIndex / resizeScale) { |
| 206 | featureCache[index - featuresOverlapIndex / resizeScale] = std::move(features); |
| 207 | } |
| 208 | }; |
| 209 | } |
| 210 | |
| 211 | template std::function<void (std::vector<int16_t>&, size_t , bool, size_t, size_t)> |
| 212 | FeatureCalc<int8_t>(TfLiteTensor* inputTensor, |
| 213 | size_t cacheSize, |
| 214 | std::function<std::vector<int8_t> (std::vector<int16_t>&)> compute); |
| 215 | |
| 216 | template std::function<void(std::vector<int16_t>&, size_t, bool, size_t, size_t)> |
| 217 | FeatureCalc<float>(TfLiteTensor *inputTensor, |
| 218 | size_t cacheSize, |
| 219 | std::function<std::vector<float>(std::vector<int16_t>&)> compute); |
| 220 | |
| 221 | std::function<void (std::vector<int16_t>&, int, bool, size_t, size_t)> |
| 222 | GetFeatureCalculator(audio::AdMelSpectrogram& melSpec, |
| 223 | TfLiteTensor* inputTensor, |
| 224 | size_t cacheSize, |
| 225 | float trainingMean); |
| 226 | |
| 227 | } /* namespace app */ |
| 228 | } /* namespace arm */ |
| 229 | |
| 230 | #endif /* AD_PROCESSING_HPP */ |