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" |
alexander | 31ae9f0 | 2022-02-10 16:15:54 +0000 | [diff] [blame^] | 30 | #include "log_macros.h" |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 31 | |
| 32 | namespace arm { |
| 33 | namespace app { |
| 34 | |
| 35 | /** |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 36 | * @brief Presents inference results using the data presentation |
| 37 | * object. |
| 38 | * @param[in] platform Reference to the hal platform object. |
| 39 | * @param[in] results Vector of classification results to be displayed. |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 40 | * @return true if successful, false otherwise. |
| 41 | **/ |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 42 | static bool PresentInferenceResult( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 43 | hal_platform& platform, |
| 44 | const std::vector<arm::app::asr::AsrResult>& results); |
| 45 | |
| 46 | /* Audio inference classification handler. */ |
| 47 | bool ClassifyAudioHandler(ApplicationContext& ctx, uint32_t clipIndex, bool runAll) |
| 48 | { |
| 49 | constexpr uint32_t dataPsnTxtInfStartX = 20; |
| 50 | constexpr uint32_t dataPsnTxtInfStartY = 40; |
| 51 | |
| 52 | auto& platform = ctx.Get<hal_platform&>("platform"); |
| 53 | platform.data_psn->clear(COLOR_BLACK); |
| 54 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 55 | auto& profiler = ctx.Get<Profiler&>("profiler"); |
| 56 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 57 | /* If the request has a valid size, set the audio index. */ |
| 58 | if (clipIndex < NUMBER_OF_FILES) { |
Éanna Ó Catháin | 8f95887 | 2021-09-15 09:32:30 +0100 | [diff] [blame] | 59 | if (!SetAppCtxIfmIdx(ctx, clipIndex,"clipIndex")) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 60 | return false; |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | /* Get model reference. */ |
| 65 | auto& model = ctx.Get<Model&>("model"); |
| 66 | if (!model.IsInited()) { |
| 67 | printf_err("Model is not initialised! Terminating processing.\n"); |
| 68 | return false; |
| 69 | } |
| 70 | |
| 71 | /* Get score threshold to be applied for the classifier (post-inference). */ |
| 72 | auto scoreThreshold = ctx.Get<float>("scoreThreshold"); |
| 73 | |
| 74 | /* Get tensors. Dimensions of the tensor should have been verified by |
| 75 | * the callee. */ |
| 76 | TfLiteTensor* inputTensor = model.GetInputTensor(0); |
| 77 | TfLiteTensor* outputTensor = model.GetOutputTensor(0); |
| 78 | const uint32_t inputRows = inputTensor->dims->data[arm::app::Wav2LetterModel::ms_inputRowsIdx]; |
| 79 | |
| 80 | /* Populate MFCC related parameters. */ |
| 81 | auto mfccParamsWinLen = ctx.Get<uint32_t>("frameLength"); |
| 82 | auto mfccParamsWinStride = ctx.Get<uint32_t>("frameStride"); |
| 83 | |
| 84 | /* Populate ASR inference context and inner lengths for input. */ |
| 85 | auto inputCtxLen = ctx.Get<uint32_t>("ctxLen"); |
| 86 | const uint32_t inputInnerLen = inputRows - (2 * inputCtxLen); |
| 87 | |
| 88 | /* Audio data stride corresponds to inputInnerLen feature vectors. */ |
| 89 | const uint32_t audioParamsWinLen = (inputRows - 1) * mfccParamsWinStride + (mfccParamsWinLen); |
| 90 | const uint32_t audioParamsWinStride = inputInnerLen * mfccParamsWinStride; |
| 91 | const float audioParamsSecondsPerSample = (1.0/audio::Wav2LetterMFCC::ms_defaultSamplingFreq); |
| 92 | |
| 93 | /* Get pre/post-processing objects. */ |
| 94 | auto& prep = ctx.Get<audio::asr::Preprocess&>("preprocess"); |
| 95 | auto& postp = ctx.Get<audio::asr::Postprocess&>("postprocess"); |
| 96 | |
| 97 | /* Set default reduction axis for post-processing. */ |
| 98 | const uint32_t reductionAxis = arm::app::Wav2LetterModel::ms_outputRowsIdx; |
| 99 | |
| 100 | /* Audio clip start index. */ |
| 101 | auto startClipIdx = ctx.Get<uint32_t>("clipIndex"); |
| 102 | |
| 103 | /* Loop to process audio clips. */ |
| 104 | do { |
Richard Burton | 9b8d67a | 2021-12-10 12:32:51 +0000 | [diff] [blame] | 105 | platform.data_psn->clear(COLOR_BLACK); |
| 106 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 107 | /* Get current audio clip index. */ |
| 108 | auto currentIndex = ctx.Get<uint32_t>("clipIndex"); |
| 109 | |
| 110 | /* Get the current audio buffer and respective size. */ |
| 111 | const int16_t* audioArr = get_audio_array(currentIndex); |
| 112 | const uint32_t audioArrSize = get_audio_array_size(currentIndex); |
| 113 | |
| 114 | if (!audioArr) { |
| 115 | printf_err("Invalid audio array pointer\n"); |
| 116 | return false; |
| 117 | } |
| 118 | |
| 119 | /* Audio clip must have enough samples to produce 1 MFCC feature. */ |
| 120 | if (audioArrSize < mfccParamsWinLen) { |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 121 | printf_err("Not enough audio samples, minimum needed is %" PRIu32 "\n", |
| 122 | mfccParamsWinLen); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 123 | return false; |
| 124 | } |
| 125 | |
| 126 | /* Initialise an audio slider. */ |
alexander | 80eecfb | 2021-07-06 19:47:59 +0100 | [diff] [blame] | 127 | auto audioDataSlider = audio::FractionalSlidingWindow<const int16_t>( |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 128 | audioArr, |
| 129 | audioArrSize, |
| 130 | audioParamsWinLen, |
| 131 | audioParamsWinStride); |
| 132 | |
| 133 | /* Declare a container for results. */ |
| 134 | std::vector<arm::app::asr::AsrResult> results; |
| 135 | |
| 136 | /* Display message on the LCD - inference running. */ |
| 137 | std::string str_inf{"Running inference... "}; |
| 138 | platform.data_psn->present_data_text( |
| 139 | str_inf.c_str(), str_inf.size(), |
| 140 | dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); |
| 141 | |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 142 | info("Running inference on audio clip %" PRIu32 " => %s\n", currentIndex, |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 143 | get_filename(currentIndex)); |
| 144 | |
| 145 | size_t inferenceWindowLen = audioParamsWinLen; |
| 146 | |
| 147 | /* Start sliding through audio clip. */ |
| 148 | while (audioDataSlider.HasNext()) { |
| 149 | |
| 150 | /* If not enough audio see how much can be sent for processing. */ |
| 151 | size_t nextStartIndex = audioDataSlider.NextWindowStartIndex(); |
| 152 | if (nextStartIndex + audioParamsWinLen > audioArrSize) { |
| 153 | inferenceWindowLen = audioArrSize - nextStartIndex; |
| 154 | } |
| 155 | |
| 156 | const int16_t* inferenceWindow = audioDataSlider.Next(); |
| 157 | |
| 158 | info("Inference %zu/%zu\n", audioDataSlider.Index() + 1, |
| 159 | static_cast<size_t>(ceilf(audioDataSlider.FractionalTotalStrides() + 1))); |
| 160 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 161 | /* Calculate MFCCs, deltas and populate the input tensor. */ |
| 162 | prep.Invoke(inferenceWindow, inferenceWindowLen, inputTensor); |
| 163 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 164 | /* Run inference over this audio clip sliding window. */ |
alexander | 27b62d9 | 2021-05-04 20:46:08 +0100 | [diff] [blame] | 165 | if (!RunInference(model, profiler)) { |
| 166 | return false; |
| 167 | } |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 168 | |
| 169 | /* Post-process. */ |
| 170 | postp.Invoke(outputTensor, reductionAxis, !audioDataSlider.HasNext()); |
| 171 | |
| 172 | /* Get results. */ |
| 173 | std::vector<ClassificationResult> classificationResult; |
| 174 | auto& classifier = ctx.Get<AsrClassifier&>("classifier"); |
| 175 | classifier.GetClassificationResults( |
| 176 | outputTensor, classificationResult, |
| 177 | ctx.Get<std::vector<std::string>&>("labels"), 1); |
| 178 | |
| 179 | results.emplace_back(asr::AsrResult(classificationResult, |
| 180 | (audioDataSlider.Index() * |
| 181 | audioParamsSecondsPerSample * |
| 182 | audioParamsWinStride), |
| 183 | audioDataSlider.Index(), scoreThreshold)); |
| 184 | |
| 185 | #if VERIFY_TEST_OUTPUT |
| 186 | arm::app::DumpTensor(outputTensor, |
| 187 | outputTensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx]); |
| 188 | #endif /* VERIFY_TEST_OUTPUT */ |
| 189 | |
| 190 | } |
| 191 | |
| 192 | /* Erase. */ |
| 193 | str_inf = std::string(str_inf.size(), ' '); |
| 194 | platform.data_psn->present_data_text( |
| 195 | str_inf.c_str(), str_inf.size(), |
| 196 | dataPsnTxtInfStartX, dataPsnTxtInfStartY, 0); |
| 197 | |
| 198 | ctx.Set<std::vector<arm::app::asr::AsrResult>>("results", results); |
| 199 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 200 | if (!PresentInferenceResult(platform, results)) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 201 | return false; |
| 202 | } |
| 203 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 204 | profiler.PrintProfilingResult(); |
| 205 | |
Éanna Ó Catháin | 8f95887 | 2021-09-15 09:32:30 +0100 | [diff] [blame] | 206 | IncrementAppCtxIfmIdx(ctx,"clipIndex"); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 207 | |
| 208 | } while (runAll && ctx.Get<uint32_t>("clipIndex") != startClipIdx); |
| 209 | |
| 210 | return true; |
| 211 | } |
| 212 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 213 | |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 214 | static bool PresentInferenceResult(hal_platform& platform, |
| 215 | const std::vector<arm::app::asr::AsrResult>& results) |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 216 | { |
| 217 | constexpr uint32_t dataPsnTxtStartX1 = 20; |
| 218 | constexpr uint32_t dataPsnTxtStartY1 = 60; |
| 219 | constexpr bool allow_multiple_lines = true; |
| 220 | |
| 221 | platform.data_psn->set_text_color(COLOR_GREEN); |
| 222 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 223 | info("Final results:\n"); |
| 224 | info("Total number of inferences: %zu\n", results.size()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 225 | /* Results from multiple inferences should be combined before processing. */ |
| 226 | std::vector<arm::app::ClassificationResult> combinedResults; |
| 227 | for (auto& result : results) { |
| 228 | combinedResults.insert(combinedResults.end(), |
| 229 | result.m_resultVec.begin(), |
| 230 | result.m_resultVec.end()); |
| 231 | } |
| 232 | |
| 233 | /* Get each inference result string using the decoder. */ |
| 234 | for (const auto & result : results) { |
| 235 | std::string infResultStr = audio::asr::DecodeOutput(result.m_resultVec); |
| 236 | |
Kshitij Sisodia | f9c19ea | 2021-05-07 16:08:14 +0100 | [diff] [blame] | 237 | info("For timestamp: %f (inference #: %" PRIu32 "); label: %s\n", |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 238 | result.m_timeStamp, result.m_inferenceNumber, |
| 239 | infResultStr.c_str()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 240 | } |
| 241 | |
| 242 | /* Get the decoded result for the combined result. */ |
| 243 | std::string finalResultStr = audio::asr::DecodeOutput(combinedResults); |
| 244 | |
| 245 | platform.data_psn->present_data_text( |
| 246 | finalResultStr.c_str(), finalResultStr.size(), |
| 247 | dataPsnTxtStartX1, dataPsnTxtStartY1, |
| 248 | allow_multiple_lines); |
| 249 | |
Isabella Gottardi | 8df12f3 | 2021-04-07 17:15:31 +0100 | [diff] [blame] | 250 | info("Complete recognition: %s\n", finalResultStr.c_str()); |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 251 | return true; |
| 252 | } |
| 253 | |
| 254 | } /* namespace app */ |
| 255 | } /* namespace arm */ |