alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2021 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 | #include "UseCaseHandler.hpp" |
| 18 | |
| 19 | #include "InputFiles.hpp" |
| 20 | #include "Classifier.hpp" |
| 21 | #include "DsCnnModel.hpp" |
| 22 | #include "hal.h" |
| 23 | #include "DsCnnMfcc.hpp" |
| 24 | #include "AudioUtils.hpp" |
| 25 | #include "UseCaseCommonUtils.hpp" |
| 26 | #include "KwsResult.hpp" |
| 27 | |
| 28 | #include <vector> |
| 29 | #include <functional> |
| 30 | |
| 31 | using KwsClassifier = arm::app::Classifier; |
| 32 | |
| 33 | namespace arm { |
| 34 | namespace app { |
| 35 | |
| 36 | /** |
| 37 | * @brief Helper function to increment current audio clip index. |
| 38 | * @param[in,out] ctx Pointer to the application context object. |
| 39 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 40 | static void IncrementAppCtxClipIdx(ApplicationContext& ctx); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 41 | |
| 42 | /** |
| 43 | * @brief Helper function to set the audio clip index. |
| 44 | * @param[in,out] ctx Pointer to the application context object. |
| 45 | * @param[in] idx Value to be set. |
| 46 | * @return true if index is set, false otherwise. |
| 47 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 48 | static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 49 | |
| 50 | /** |
| 51 | * @brief Presents inference results using the data presentation |
| 52 | * object. |
| 53 | * @param[in] platform Reference to the hal platform object. |
| 54 | * @param[in] results Vector of classification results to be displayed. |
| 55 | * @param[in] infTimeMs Inference time in milliseconds, if available, |
| 56 | * otherwise, this can be passed in as 0. |
| 57 | * @return true if successful, false otherwise. |
| 58 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 59 | static bool PresentInferenceResult(hal_platform& platform, |
| 60 | const std::vector<arm::app::kws::KwsResult>& results); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 61 | |
| 62 | /** |
| 63 | * @brief Returns a function to perform feature calculation and populates input tensor data with |
| 64 | * MFCC data. |
| 65 | * |
| 66 | * Input tensor data type check is performed to choose correct MFCC feature data type. |
| 67 | * If tensor has an integer data type then original features are quantised. |
| 68 | * |
| 69 | * Warning: MFCC calculator provided as input must have the same life scope as returned function. |
| 70 | * |
| 71 | * @param[in] mfcc MFCC feature calculator. |
| 72 | * @param[in,out] inputTensor Input tensor pointer to store calculated features. |
| 73 | * @param[in] cacheSize Size of the feature vectors cache (number of feature vectors). |
| 74 | * @return Function to be called providing audio sample and sliding window index. |
| 75 | */ |
| 76 | static std::function<void (std::vector<int16_t>&, int, bool, size_t)> |
| 77 | GetFeatureCalculator(audio::DsCnnMFCC& mfcc, |
| 78 | TfLiteTensor* inputTensor, |
| 79 | size_t cacheSize); |
| 80 | |
| 81 | /* Audio inference handler. */ |
| 82 | bool ClassifyAudioHandler(ApplicationContext& ctx, uint32_t clipIndex, bool runAll) |
| 83 | { |
| 84 | auto& platform = ctx.Get<hal_platform&>("platform"); |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 85 | auto& profiler = ctx.Get<Profiler&>("profiler"); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 86 | |
| 87 | constexpr uint32_t dataPsnTxtInfStartX = 20; |
| 88 | constexpr uint32_t dataPsnTxtInfStartY = 40; |
| 89 | constexpr int minTensorDims = static_cast<int>( |
| 90 | (arm::app::DsCnnModel::ms_inputRowsIdx > arm::app::DsCnnModel::ms_inputColsIdx)? |
| 91 | arm::app::DsCnnModel::ms_inputRowsIdx : arm::app::DsCnnModel::ms_inputColsIdx); |
| 92 | |
| 93 | platform.data_psn->clear(COLOR_BLACK); |
| 94 | |
| 95 | auto& model = ctx.Get<Model&>("model"); |
| 96 | |
| 97 | /* If the request has a valid size, set the audio index. */ |
| 98 | if (clipIndex < NUMBER_OF_FILES) { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 99 | if (!SetAppCtxClipIdx(ctx, clipIndex)) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 100 | return false; |
| 101 | } |
| 102 | } |
| 103 | if (!model.IsInited()) { |
| 104 | printf_err("Model is not initialised! Terminating processing.\n"); |
| 105 | return false; |
| 106 | } |
| 107 | |
| 108 | const auto frameLength = ctx.Get<int>("frameLength"); |
| 109 | const auto frameStride = ctx.Get<int>("frameStride"); |
| 110 | const auto scoreThreshold = ctx.Get<float>("scoreThreshold"); |
| 111 | auto startClipIdx = ctx.Get<uint32_t>("clipIndex"); |
| 112 | |
| 113 | TfLiteTensor* outputTensor = model.GetOutputTensor(0); |
| 114 | TfLiteTensor* inputTensor = model.GetInputTensor(0); |
| 115 | |
| 116 | if (!inputTensor->dims) { |
| 117 | printf_err("Invalid input tensor dims\n"); |
| 118 | return false; |
| 119 | } else if (inputTensor->dims->size < minTensorDims) { |
| 120 | printf_err("Input tensor dimension should be >= %d\n", minTensorDims); |
| 121 | return false; |
| 122 | } |
| 123 | |
| 124 | TfLiteIntArray* inputShape = model.GetInputShape(0); |
| 125 | const uint32_t kNumCols = inputShape->data[arm::app::DsCnnModel::ms_inputColsIdx]; |
| 126 | const uint32_t kNumRows = inputShape->data[arm::app::DsCnnModel::ms_inputRowsIdx]; |
| 127 | |
| 128 | audio::DsCnnMFCC mfcc = audio::DsCnnMFCC(kNumCols, frameLength); |
| 129 | mfcc.Init(); |
| 130 | |
| 131 | /* Deduce the data length required for 1 inference from the network parameters. */ |
| 132 | auto audioDataWindowSize = kNumRows * frameStride + (frameLength - frameStride); |
| 133 | auto mfccWindowSize = frameLength; |
| 134 | auto mfccWindowStride = frameStride; |
| 135 | |
| 136 | /* We choose to move by half the window size => for a 1 second window size |
| 137 | * there is an overlap of 0.5 seconds. */ |
| 138 | auto audioDataStride = audioDataWindowSize / 2; |
| 139 | |
| 140 | /* To have the previously calculated features re-usable, stride must be multiple |
| 141 | * of MFCC features window stride. */ |
| 142 | if (0 != audioDataStride % mfccWindowStride) { |
| 143 | |
| 144 | /* Reduce the stride. */ |
| 145 | audioDataStride -= audioDataStride % mfccWindowStride; |
| 146 | } |
| 147 | |
| 148 | auto nMfccVectorsInAudioStride = audioDataStride/mfccWindowStride; |
| 149 | |
| 150 | /* We expect to be sampling 1 second worth of data at a time. |
| 151 | * NOTE: This is only used for time stamp calculation. */ |
| 152 | const float secondsPerSample = 1.0/audio::DsCnnMFCC::ms_defaultSamplingFreq; |
| 153 | |
| 154 | do { |
| 155 | auto currentIndex = ctx.Get<uint32_t>("clipIndex"); |
| 156 | |
| 157 | /* Creating a mfcc features sliding window for the data required for 1 inference. */ |
| 158 | auto audioMFCCWindowSlider = audio::SlidingWindow<const int16_t>( |
| 159 | get_audio_array(currentIndex), |
| 160 | audioDataWindowSize, mfccWindowSize, |
| 161 | mfccWindowStride); |
| 162 | |
| 163 | /* Creating a sliding window through the whole audio clip. */ |
| 164 | auto audioDataSlider = audio::SlidingWindow<const int16_t>( |
| 165 | get_audio_array(currentIndex), |
| 166 | get_audio_array_size(currentIndex), |
| 167 | audioDataWindowSize, audioDataStride); |
| 168 | |
| 169 | /* Calculate number of the feature vectors in the window overlap region. |
| 170 | * These feature vectors will be reused.*/ |
| 171 | auto numberOfReusedFeatureVectors = audioMFCCWindowSlider.TotalStrides() + 1 |
| 172 | - nMfccVectorsInAudioStride; |
| 173 | |
| 174 | /* Construct feature calculation function. */ |
| 175 | auto mfccFeatureCalc = GetFeatureCalculator(mfcc, inputTensor, |
| 176 | numberOfReusedFeatureVectors); |
| 177 | |
| 178 | if (!mfccFeatureCalc){ |
| 179 | return false; |
| 180 | } |
| 181 | |
| 182 | /* Declare a container for results. */ |
| 183 | std::vector<arm::app::kws::KwsResult> results; |
| 184 | |
| 185 | /* Display message on the LCD - inference running. */ |
| 186 | std::string str_inf{"Running inference... "}; |
| 187 | platform.data_psn->present_data_text( |
| 188 | str_inf.c_str(), str_inf.size(), |
| 189 | dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); |
| 190 | info("Running inference on audio clip %u => %s\n", currentIndex, |
| 191 | get_filename(currentIndex)); |
| 192 | |
| 193 | /* Start sliding through audio clip. */ |
| 194 | while (audioDataSlider.HasNext()) { |
| 195 | const int16_t *inferenceWindow = audioDataSlider.Next(); |
| 196 | |
| 197 | /* We moved to the next window - set the features sliding to the new address. */ |
| 198 | audioMFCCWindowSlider.Reset(inferenceWindow); |
| 199 | |
| 200 | /* The first window does not have cache ready. */ |
| 201 | bool useCache = audioDataSlider.Index() > 0 && numberOfReusedFeatureVectors > 0; |
| 202 | |
| 203 | /* Start calculating features inside one audio sliding window. */ |
| 204 | while (audioMFCCWindowSlider.HasNext()) { |
| 205 | const int16_t *mfccWindow = audioMFCCWindowSlider.Next(); |
| 206 | std::vector<int16_t> mfccAudioData = std::vector<int16_t>(mfccWindow, |
| 207 | mfccWindow + mfccWindowSize); |
| 208 | /* Compute features for this window and write them to input tensor. */ |
| 209 | mfccFeatureCalc(mfccAudioData, |
| 210 | audioMFCCWindowSlider.Index(), |
| 211 | useCache, |
| 212 | nMfccVectorsInAudioStride); |
| 213 | } |
| 214 | |
| 215 | info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, |
| 216 | audioDataSlider.TotalStrides() + 1); |
| 217 | |
| 218 | /* Run inference over this audio clip sliding window. */ |
alexander | 27b62d9 | 2021-05-04 20:46:08 +0100 | [diff] [blame^] | 219 | if (!RunInference(model, profiler)) { |
| 220 | return false; |
| 221 | } |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 222 | |
| 223 | std::vector<ClassificationResult> classificationResult; |
| 224 | auto& classifier = ctx.Get<KwsClassifier&>("classifier"); |
| 225 | classifier.GetClassificationResults(outputTensor, classificationResult, |
| 226 | ctx.Get<std::vector<std::string>&>("labels"), 1); |
| 227 | |
| 228 | results.emplace_back(kws::KwsResult(classificationResult, |
| 229 | audioDataSlider.Index() * secondsPerSample * audioDataStride, |
| 230 | audioDataSlider.Index(), scoreThreshold)); |
| 231 | |
| 232 | #if VERIFY_TEST_OUTPUT |
| 233 | arm::app::DumpTensor(outputTensor); |
| 234 | #endif /* VERIFY_TEST_OUTPUT */ |
| 235 | } /* while (audioDataSlider.HasNext()) */ |
| 236 | |
| 237 | /* Erase. */ |
| 238 | str_inf = std::string(str_inf.size(), ' '); |
| 239 | platform.data_psn->present_data_text( |
| 240 | str_inf.c_str(), str_inf.size(), |
| 241 | dataPsnTxtInfStartX, dataPsnTxtInfStartY, false); |
| 242 | |
| 243 | ctx.Set<std::vector<arm::app::kws::KwsResult>>("results", results); |
| 244 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 245 | if (!PresentInferenceResult(platform, results)) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 246 | return false; |
| 247 | } |
| 248 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 249 | profiler.PrintProfilingResult(); |
| 250 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 251 | IncrementAppCtxClipIdx(ctx); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 252 | |
| 253 | } while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx); |
| 254 | |
| 255 | return true; |
| 256 | } |
| 257 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 258 | static void IncrementAppCtxClipIdx(ApplicationContext& ctx) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 259 | { |
| 260 | auto curAudioIdx = ctx.Get<uint32_t>("clipIndex"); |
| 261 | |
| 262 | if (curAudioIdx + 1 >= NUMBER_OF_FILES) { |
| 263 | ctx.Set<uint32_t>("clipIndex", 0); |
| 264 | return; |
| 265 | } |
| 266 | ++curAudioIdx; |
| 267 | ctx.Set<uint32_t>("clipIndex", curAudioIdx); |
| 268 | } |
| 269 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 270 | static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 271 | { |
| 272 | if (idx >= NUMBER_OF_FILES) { |
| 273 | printf_err("Invalid idx %u (expected less than %u)\n", |
| 274 | idx, NUMBER_OF_FILES); |
| 275 | return false; |
| 276 | } |
| 277 | ctx.Set<uint32_t>("clipIndex", idx); |
| 278 | return true; |
| 279 | } |
| 280 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 281 | static bool PresentInferenceResult(hal_platform& platform, |
| 282 | const std::vector<arm::app::kws::KwsResult>& results) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 283 | { |
| 284 | constexpr uint32_t dataPsnTxtStartX1 = 20; |
| 285 | constexpr uint32_t dataPsnTxtStartY1 = 30; |
| 286 | constexpr uint32_t dataPsnTxtYIncr = 16; /* Row index increment. */ |
| 287 | |
| 288 | platform.data_psn->set_text_color(COLOR_GREEN); |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 289 | info("Final results:\n"); |
| 290 | info("Total number of inferences: %zu\n", results.size()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 291 | |
| 292 | /* Display each result */ |
| 293 | uint32_t rowIdx1 = dataPsnTxtStartY1 + 2 * dataPsnTxtYIncr; |
| 294 | |
| 295 | for (uint32_t i = 0; i < results.size(); ++i) { |
| 296 | |
| 297 | std::string topKeyword{"<none>"}; |
| 298 | float score = 0.f; |
| 299 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 300 | if (!results[i].m_resultVec.empty()) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 301 | topKeyword = results[i].m_resultVec[0].m_label; |
| 302 | score = results[i].m_resultVec[0].m_normalisedVal; |
| 303 | } |
| 304 | |
| 305 | std::string resultStr = |
| 306 | std::string{"@"} + std::to_string(results[i].m_timeStamp) + |
| 307 | std::string{"s: "} + topKeyword + std::string{" ("} + |
| 308 | std::to_string(static_cast<int>(score * 100)) + std::string{"%)"}; |
| 309 | |
| 310 | platform.data_psn->present_data_text( |
| 311 | resultStr.c_str(), resultStr.size(), |
| 312 | dataPsnTxtStartX1, rowIdx1, false); |
| 313 | rowIdx1 += dataPsnTxtYIncr; |
| 314 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 315 | if (results[i].m_resultVec.empty()) { |
| 316 | info("For timestamp: %f (inference #: %u); label: %s; threshold: %f\n", |
| 317 | results[i].m_timeStamp, results[i].m_inferenceNumber, |
| 318 | topKeyword.c_str(), |
| 319 | results[i].m_threshold); |
| 320 | } else { |
| 321 | for (uint32_t j = 0; j < results[i].m_resultVec.size(); ++j) { |
| 322 | info("For timestamp: %f (inference #: %u); label: %s, score: %f; threshold: %f\n", |
| 323 | results[i].m_timeStamp, |
| 324 | results[i].m_inferenceNumber, |
| 325 | results[i].m_resultVec[j].m_label.c_str(), |
| 326 | results[i].m_resultVec[j].m_normalisedVal, |
| 327 | results[i].m_threshold); |
| 328 | } |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 329 | } |
| 330 | } |
| 331 | |
| 332 | return true; |
| 333 | } |
| 334 | |
| 335 | /** |
| 336 | * @brief Generic feature calculator factory. |
| 337 | * |
| 338 | * Returns lambda function to compute features using features cache. |
| 339 | * Real features math is done by a lambda function provided as a parameter. |
| 340 | * Features are written to input tensor memory. |
| 341 | * |
| 342 | * @tparam T Feature vector type. |
| 343 | * @param inputTensor Model input tensor pointer. |
| 344 | * @param cacheSize Number of feature vectors to cache. Defined by the sliding window overlap. |
| 345 | * @param compute Features calculator function. |
| 346 | * @return Lambda function to compute features. |
| 347 | */ |
| 348 | template<class T> |
| 349 | std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 350 | FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, |
| 351 | std::function<std::vector<T> (std::vector<int16_t>& )> compute) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 352 | { |
| 353 | /* Feature cache to be captured by lambda function. */ |
| 354 | static std::vector<std::vector<T>> featureCache = std::vector<std::vector<T>>(cacheSize); |
| 355 | |
| 356 | return [=](std::vector<int16_t>& audioDataWindow, |
| 357 | size_t index, |
| 358 | bool useCache, |
| 359 | size_t featuresOverlapIndex) |
| 360 | { |
| 361 | T *tensorData = tflite::GetTensorData<T>(inputTensor); |
| 362 | std::vector<T> features; |
| 363 | |
| 364 | /* Reuse features from cache if cache is ready and sliding windows overlap. |
| 365 | * Overlap is in the beginning of sliding window with a size of a feature cache. */ |
| 366 | if (useCache && index < featureCache.size()) { |
| 367 | features = std::move(featureCache[index]); |
| 368 | } else { |
| 369 | features = std::move(compute(audioDataWindow)); |
| 370 | } |
| 371 | auto size = features.size(); |
| 372 | auto sizeBytes = sizeof(T) * size; |
| 373 | std::memcpy(tensorData + (index * size), features.data(), sizeBytes); |
| 374 | |
| 375 | /* Start renewing cache as soon iteration goes out of the windows overlap. */ |
| 376 | if (index >= featuresOverlapIndex) { |
| 377 | featureCache[index - featuresOverlapIndex] = std::move(features); |
| 378 | } |
| 379 | }; |
| 380 | } |
| 381 | |
| 382 | template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 383 | FeatureCalc<int8_t>(TfLiteTensor* inputTensor, |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 384 | size_t cacheSize, |
| 385 | std::function<std::vector<int8_t> (std::vector<int16_t>& )> compute); |
| 386 | |
| 387 | template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 388 | FeatureCalc<uint8_t>(TfLiteTensor* inputTensor, |
| 389 | size_t cacheSize, |
| 390 | std::function<std::vector<uint8_t> (std::vector<int16_t>& )> compute); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 391 | |
| 392 | template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 393 | FeatureCalc<int16_t>(TfLiteTensor* inputTensor, |
| 394 | size_t cacheSize, |
| 395 | std::function<std::vector<int16_t> (std::vector<int16_t>& )> compute); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 396 | |
| 397 | template std::function<void(std::vector<int16_t>&, size_t, bool, size_t)> |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 398 | FeatureCalc<float>(TfLiteTensor* inputTensor, |
| 399 | size_t cacheSize, |
| 400 | std::function<std::vector<float>(std::vector<int16_t>&)> compute); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 401 | |
| 402 | |
| 403 | static std::function<void (std::vector<int16_t>&, int, bool, size_t)> |
| 404 | GetFeatureCalculator(audio::DsCnnMFCC& mfcc, TfLiteTensor* inputTensor, size_t cacheSize) |
| 405 | { |
| 406 | std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> mfccFeatureCalc; |
| 407 | |
| 408 | TfLiteQuantization quant = inputTensor->quantization; |
| 409 | |
| 410 | if (kTfLiteAffineQuantization == quant.type) { |
| 411 | |
| 412 | auto *quantParams = (TfLiteAffineQuantization *) quant.params; |
| 413 | const float quantScale = quantParams->scale->data[0]; |
| 414 | const int quantOffset = quantParams->zero_point->data[0]; |
| 415 | |
| 416 | switch (inputTensor->type) { |
| 417 | case kTfLiteInt8: { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 418 | mfccFeatureCalc = FeatureCalc<int8_t>(inputTensor, |
| 419 | cacheSize, |
| 420 | [=, &mfcc](std::vector<int16_t>& audioDataWindow) { |
| 421 | return mfcc.MfccComputeQuant<int8_t>(audioDataWindow, |
| 422 | quantScale, |
| 423 | quantOffset); |
| 424 | } |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 425 | ); |
| 426 | break; |
| 427 | } |
| 428 | case kTfLiteUInt8: { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 429 | mfccFeatureCalc = FeatureCalc<uint8_t>(inputTensor, |
| 430 | cacheSize, |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 431 | [=, &mfcc](std::vector<int16_t>& audioDataWindow) { |
| 432 | return mfcc.MfccComputeQuant<uint8_t>(audioDataWindow, |
| 433 | quantScale, |
| 434 | quantOffset); |
| 435 | } |
| 436 | ); |
| 437 | break; |
| 438 | } |
| 439 | case kTfLiteInt16: { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 440 | mfccFeatureCalc = FeatureCalc<int16_t>(inputTensor, |
| 441 | cacheSize, |
| 442 | [=, &mfcc](std::vector<int16_t>& audioDataWindow) { |
| 443 | return mfcc.MfccComputeQuant<int16_t>(audioDataWindow, |
| 444 | quantScale, |
| 445 | quantOffset); |
| 446 | } |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 447 | ); |
| 448 | break; |
| 449 | } |
| 450 | default: |
| 451 | printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); |
| 452 | } |
| 453 | |
| 454 | |
| 455 | } else { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 456 | mfccFeatureCalc = mfccFeatureCalc = FeatureCalc<float>(inputTensor, |
| 457 | cacheSize, |
| 458 | [&mfcc](std::vector<int16_t>& audioDataWindow) { |
| 459 | return mfcc.MfccCompute(audioDataWindow); |
| 460 | }); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 461 | } |
| 462 | return mfccFeatureCalc; |
| 463 | } |
| 464 | |
| 465 | } /* namespace app */ |
| 466 | } /* namespace arm */ |