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 "AsrClassifier.hpp" |
| 21 | #include "Wav2LetterModel.hpp" |
| 22 | #include "hal.h" |
| 23 | #include "Wav2LetterMfcc.hpp" |
| 24 | #include "AudioUtils.hpp" |
| 25 | #include "UseCaseCommonUtils.hpp" |
| 26 | #include "AsrResult.hpp" |
| 27 | #include "Wav2LetterPreprocess.hpp" |
| 28 | #include "Wav2LetterPostprocess.hpp" |
| 29 | #include "OutputDecode.hpp" |
| 30 | |
| 31 | namespace arm { |
| 32 | namespace app { |
| 33 | |
| 34 | /** |
| 35 | * @brief Helper function to increment current audio clip index. |
| 36 | * @param[in,out] ctx Pointer to the application context object. |
| 37 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 38 | static void IncrementAppCtxClipIdx(ApplicationContext& ctx); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 39 | |
| 40 | /** |
| 41 | * @brief Helper function to set the audio clip index. |
| 42 | * @param[in,out] ctx Pointer to the application context object. |
| 43 | * @param[in] idx Value to be set. |
| 44 | * @return true if index is set, false otherwise. |
| 45 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 46 | static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 47 | |
| 48 | /** |
| 49 | * @brief Presents inference results using the data presentation |
| 50 | * object. |
| 51 | * @param[in] platform Reference to the hal platform object. |
| 52 | * @param[in] results Vector of classification results to be displayed. |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 53 | * @return true if successful, false otherwise. |
| 54 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 55 | static bool PresentInferenceResult( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 56 | hal_platform& platform, |
| 57 | const std::vector<arm::app::asr::AsrResult>& results); |
| 58 | |
| 59 | /* Audio inference classification handler. */ |
| 60 | bool ClassifyAudioHandler(ApplicationContext& ctx, uint32_t clipIndex, bool runAll) |
| 61 | { |
| 62 | constexpr uint32_t dataPsnTxtInfStartX = 20; |
| 63 | constexpr uint32_t dataPsnTxtInfStartY = 40; |
| 64 | |
| 65 | auto& platform = ctx.Get<hal_platform&>("platform"); |
| 66 | platform.data_psn->clear(COLOR_BLACK); |
| 67 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 68 | auto& profiler = ctx.Get<Profiler&>("profiler"); |
| 69 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 70 | /* If the request has a valid size, set the audio index. */ |
| 71 | if (clipIndex < NUMBER_OF_FILES) { |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 72 | if (!SetAppCtxClipIdx(ctx, clipIndex)) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 73 | return false; |
| 74 | } |
| 75 | } |
| 76 | |
| 77 | /* Get model reference. */ |
| 78 | auto& model = ctx.Get<Model&>("model"); |
| 79 | if (!model.IsInited()) { |
| 80 | printf_err("Model is not initialised! Terminating processing.\n"); |
| 81 | return false; |
| 82 | } |
| 83 | |
| 84 | /* Get score threshold to be applied for the classifier (post-inference). */ |
| 85 | auto scoreThreshold = ctx.Get<float>("scoreThreshold"); |
| 86 | |
| 87 | /* Get tensors. Dimensions of the tensor should have been verified by |
| 88 | * the callee. */ |
| 89 | TfLiteTensor* inputTensor = model.GetInputTensor(0); |
| 90 | TfLiteTensor* outputTensor = model.GetOutputTensor(0); |
| 91 | const uint32_t inputRows = inputTensor->dims->data[arm::app::Wav2LetterModel::ms_inputRowsIdx]; |
| 92 | |
| 93 | /* Populate MFCC related parameters. */ |
| 94 | auto mfccParamsWinLen = ctx.Get<uint32_t>("frameLength"); |
| 95 | auto mfccParamsWinStride = ctx.Get<uint32_t>("frameStride"); |
| 96 | |
| 97 | /* Populate ASR inference context and inner lengths for input. */ |
| 98 | auto inputCtxLen = ctx.Get<uint32_t>("ctxLen"); |
| 99 | const uint32_t inputInnerLen = inputRows - (2 * inputCtxLen); |
| 100 | |
| 101 | /* Audio data stride corresponds to inputInnerLen feature vectors. */ |
| 102 | const uint32_t audioParamsWinLen = (inputRows - 1) * mfccParamsWinStride + (mfccParamsWinLen); |
| 103 | const uint32_t audioParamsWinStride = inputInnerLen * mfccParamsWinStride; |
| 104 | const float audioParamsSecondsPerSample = (1.0/audio::Wav2LetterMFCC::ms_defaultSamplingFreq); |
| 105 | |
| 106 | /* Get pre/post-processing objects. */ |
| 107 | auto& prep = ctx.Get<audio::asr::Preprocess&>("preprocess"); |
| 108 | auto& postp = ctx.Get<audio::asr::Postprocess&>("postprocess"); |
| 109 | |
| 110 | /* Set default reduction axis for post-processing. */ |
| 111 | const uint32_t reductionAxis = arm::app::Wav2LetterModel::ms_outputRowsIdx; |
| 112 | |
| 113 | /* Audio clip start index. */ |
| 114 | auto startClipIdx = ctx.Get<uint32_t>("clipIndex"); |
| 115 | |
| 116 | /* Loop to process audio clips. */ |
| 117 | do { |
| 118 | /* Get current audio clip index. */ |
| 119 | auto currentIndex = ctx.Get<uint32_t>("clipIndex"); |
| 120 | |
| 121 | /* Get the current audio buffer and respective size. */ |
| 122 | const int16_t* audioArr = get_audio_array(currentIndex); |
| 123 | const uint32_t audioArrSize = get_audio_array_size(currentIndex); |
| 124 | |
| 125 | if (!audioArr) { |
| 126 | printf_err("Invalid audio array pointer\n"); |
| 127 | return false; |
| 128 | } |
| 129 | |
| 130 | /* Audio clip must have enough samples to produce 1 MFCC feature. */ |
| 131 | if (audioArrSize < mfccParamsWinLen) { |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 132 | printf_err("Not enough audio samples, minimum needed is %" PRIu32 "\n", |
| 133 | mfccParamsWinLen); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 134 | return false; |
| 135 | } |
| 136 | |
| 137 | /* Initialise an audio slider. */ |
alexander | 80eecfb | 2021-07-06 19:47:59 +0100 | [diff] [blame] | 138 | auto audioDataSlider = audio::FractionalSlidingWindow<const int16_t>( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 139 | audioArr, |
| 140 | audioArrSize, |
| 141 | audioParamsWinLen, |
| 142 | audioParamsWinStride); |
| 143 | |
| 144 | /* Declare a container for results. */ |
| 145 | std::vector<arm::app::asr::AsrResult> results; |
| 146 | |
| 147 | /* Display message on the LCD - inference running. */ |
| 148 | std::string str_inf{"Running inference... "}; |
| 149 | platform.data_psn->present_data_text( |
| 150 | str_inf.c_str(), str_inf.size(), |
| 151 | dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); |
| 152 | |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 153 | info("Running inference on audio clip %" PRIu32 " => %s\n", currentIndex, |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 154 | get_filename(currentIndex)); |
| 155 | |
| 156 | size_t inferenceWindowLen = audioParamsWinLen; |
| 157 | |
| 158 | /* Start sliding through audio clip. */ |
| 159 | while (audioDataSlider.HasNext()) { |
| 160 | |
| 161 | /* If not enough audio see how much can be sent for processing. */ |
| 162 | size_t nextStartIndex = audioDataSlider.NextWindowStartIndex(); |
| 163 | if (nextStartIndex + audioParamsWinLen > audioArrSize) { |
| 164 | inferenceWindowLen = audioArrSize - nextStartIndex; |
| 165 | } |
| 166 | |
| 167 | const int16_t* inferenceWindow = audioDataSlider.Next(); |
| 168 | |
| 169 | info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, |
| 170 | static_cast<size_t>(ceilf(audioDataSlider.FractionalTotalStrides() + 1))); |
| 171 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 172 | /* Calculate MFCCs, deltas and populate the input tensor. */ |
| 173 | prep.Invoke(inferenceWindow, inferenceWindowLen, inputTensor); |
| 174 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 175 | /* Run inference over this audio clip sliding window. */ |
alexander | 27b62d9 | 2021-05-04 20:46:08 +0100 | [diff] [blame] | 176 | if (!RunInference(model, profiler)) { |
| 177 | return false; |
| 178 | } |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 179 | |
| 180 | /* Post-process. */ |
| 181 | postp.Invoke(outputTensor, reductionAxis, !audioDataSlider.HasNext()); |
| 182 | |
| 183 | /* Get results. */ |
| 184 | std::vector<ClassificationResult> classificationResult; |
| 185 | auto& classifier = ctx.Get<AsrClassifier&>("classifier"); |
| 186 | classifier.GetClassificationResults( |
| 187 | outputTensor, classificationResult, |
| 188 | ctx.Get<std::vector<std::string>&>("labels"), 1); |
| 189 | |
| 190 | results.emplace_back(asr::AsrResult(classificationResult, |
| 191 | (audioDataSlider.Index() * |
| 192 | audioParamsSecondsPerSample * |
| 193 | audioParamsWinStride), |
| 194 | audioDataSlider.Index(), scoreThreshold)); |
| 195 | |
| 196 | #if VERIFY_TEST_OUTPUT |
| 197 | arm::app::DumpTensor(outputTensor, |
| 198 | outputTensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx]); |
| 199 | #endif /* VERIFY_TEST_OUTPUT */ |
| 200 | |
| 201 | } |
| 202 | |
| 203 | /* Erase. */ |
| 204 | str_inf = std::string(str_inf.size(), ' '); |
| 205 | platform.data_psn->present_data_text( |
| 206 | str_inf.c_str(), str_inf.size(), |
| 207 | dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); |
| 208 | |
| 209 | ctx.Set<std::vector<arm::app::asr::AsrResult>>("results", results); |
| 210 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 211 | if (!PresentInferenceResult(platform, results)) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 212 | return false; |
| 213 | } |
| 214 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 215 | profiler.PrintProfilingResult(); |
| 216 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 217 | IncrementAppCtxClipIdx(ctx); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 218 | |
| 219 | } while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx); |
| 220 | |
| 221 | return true; |
| 222 | } |
| 223 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 224 | static void IncrementAppCtxClipIdx(ApplicationContext& ctx) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 225 | { |
| 226 | auto curAudioIdx = ctx.Get<uint32_t>("clipIndex"); |
| 227 | |
| 228 | if (curAudioIdx + 1 >= NUMBER_OF_FILES) { |
| 229 | ctx.Set<uint32_t>("clipIndex", 0); |
| 230 | return; |
| 231 | } |
| 232 | ++curAudioIdx; |
| 233 | ctx.Set<uint32_t>("clipIndex", curAudioIdx); |
| 234 | } |
| 235 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 236 | static bool SetAppCtxClipIdx(ApplicationContext& ctx, uint32_t idx) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 237 | { |
| 238 | if (idx >= NUMBER_OF_FILES) { |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 239 | printf_err("Invalid idx %" PRIu32 " (expected less than %u)\n", |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 240 | idx, NUMBER_OF_FILES); |
| 241 | return false; |
| 242 | } |
| 243 | |
| 244 | ctx.Set<uint32_t>("clipIndex", idx); |
| 245 | return true; |
| 246 | } |
| 247 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 248 | static bool PresentInferenceResult(hal_platform& platform, |
| 249 | const std::vector<arm::app::asr::AsrResult>& results) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 250 | { |
| 251 | constexpr uint32_t dataPsnTxtStartX1 = 20; |
| 252 | constexpr uint32_t dataPsnTxtStartY1 = 60; |
| 253 | constexpr bool allow_multiple_lines = true; |
| 254 | |
| 255 | platform.data_psn->set_text_color(COLOR_GREEN); |
| 256 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 257 | info("Final results:\n"); |
| 258 | info("Total number of inferences: %zu\n", results.size()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 259 | /* Results from multiple inferences should be combined before processing. */ |
| 260 | std::vector<arm::app::ClassificationResult> combinedResults; |
| 261 | for (auto& result : results) { |
| 262 | combinedResults.insert(combinedResults.end(), |
| 263 | result.m_resultVec.begin(), |
| 264 | result.m_resultVec.end()); |
| 265 | } |
| 266 | |
| 267 | /* Get each inference result string using the decoder. */ |
| 268 | for (const auto & result : results) { |
| 269 | std::string infResultStr = audio::asr::DecodeOutput(result.m_resultVec); |
| 270 | |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 271 | info("For timestamp: %f (inference #: %" PRIu32 "); label: %s\n", |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 272 | result.m_timeStamp, result.m_inferenceNumber, |
| 273 | infResultStr.c_str()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 274 | } |
| 275 | |
| 276 | /* Get the decoded result for the combined result. */ |
| 277 | std::string finalResultStr = audio::asr::DecodeOutput(combinedResults); |
| 278 | |
| 279 | platform.data_psn->present_data_text( |
| 280 | finalResultStr.c_str(), finalResultStr.size(), |
| 281 | dataPsnTxtStartX1, dataPsnTxtStartY1, |
| 282 | allow_multiple_lines); |
| 283 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 284 | info("Complete recognition: %s\n", finalResultStr.c_str()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 285 | return true; |
| 286 | } |
| 287 | |
| 288 | } /* namespace app */ |
| 289 | } /* namespace arm */ |