diff options
Diffstat (limited to 'source/use_case/kws_asr')
22 files changed, 48 insertions, 2048 deletions
diff --git a/source/use_case/kws_asr/include/AsrClassifier.hpp b/source/use_case/kws_asr/include/AsrClassifier.hpp deleted file mode 100644 index 6ab9685..0000000 --- a/source/use_case/kws_asr/include/AsrClassifier.hpp +++ /dev/null @@ -1,66 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef ASR_CLASSIFIER_HPP -#define ASR_CLASSIFIER_HPP - -#include "Classifier.hpp" - -namespace arm { -namespace app { - - class AsrClassifier : public Classifier { - public: - /** - * @brief Gets the top N classification results from the - * output vector. - * @param[in] outputTensor Inference output tensor from an NN model. - * @param[out] vecResults A vector of classification results - * populated by this function. - * @param[in] labels Labels vector to match classified classes - * @param[in] topNCount Number of top classifications to pick. - * @param[in] use_softmax Whether softmax scaling should be applied to model output. - * @return true if successful, false otherwise. - **/ - bool GetClassificationResults( - TfLiteTensor* outputTensor, - std::vector<ClassificationResult>& vecResults, - const std::vector <std::string>& labels, uint32_t topNCount, - bool use_softmax = false) override; - - private: - - /** - * @brief Utility function that gets the top 1 classification results from the - * output tensor (vector of vector). - * @param[in] tensor Inference output tensor from an NN model. - * @param[out] vecResults A vector of classification results - * populated by this function. - * @param[in] labels Labels vector to match classified classes. - * @param[in] scale Quantization scale. - * @param[in] zeroPoint Quantization zero point. - * @return true if successful, false otherwise. - **/ - template<typename T> - bool GetTopResults(TfLiteTensor* tensor, - std::vector<ClassificationResult>& vecResults, - const std::vector <std::string>& labels, double scale, double zeroPoint); - }; - -} /* namespace app */ -} /* namespace arm */ - -#endif /* ASR_CLASSIFIER_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/include/AsrResult.hpp b/source/use_case/kws_asr/include/AsrResult.hpp deleted file mode 100644 index 25fa9e8..0000000 --- a/source/use_case/kws_asr/include/AsrResult.hpp +++ /dev/null @@ -1,63 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef ASR_RESULT_HPP -#define ASR_RESULT_HPP - -#include "ClassificationResult.hpp" - -#include <vector> - -namespace arm { -namespace app { -namespace asr { - - using ResultVec = std::vector<arm::app::ClassificationResult>; - - /* Structure for holding asr result. */ - class AsrResult { - - public: - ResultVec m_resultVec; /* Container for "thresholded" classification results. */ - float m_timeStamp; /* Audio timestamp for this result. */ - uint32_t m_inferenceNumber; /* Corresponding inference number. */ - float m_threshold; /* Threshold value for `m_resultVec` */ - - AsrResult() = delete; - AsrResult(ResultVec& resultVec, - const float timestamp, - const uint32_t inferenceIdx, - const float scoreThreshold) { - - this->m_threshold = scoreThreshold; - this->m_timeStamp = timestamp; - this->m_inferenceNumber = inferenceIdx; - - this->m_resultVec = ResultVec(); - for (auto& i : resultVec) { - if (i.m_normalisedVal >= this->m_threshold) { - this->m_resultVec.emplace_back(i); - } - } - } - ~AsrResult() = default; - }; - -} /* namespace asr */ -} /* namespace app */ -} /* namespace arm */ - -#endif /* ASR_RESULT_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/include/KwsProcessing.hpp b/source/use_case/kws_asr/include/KwsProcessing.hpp deleted file mode 100644 index d3de3b3..0000000 --- a/source/use_case/kws_asr/include/KwsProcessing.hpp +++ /dev/null @@ -1,138 +0,0 @@ -/* - * Copyright (c) 2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_PROCESSING_HPP -#define KWS_PROCESSING_HPP - -#include <AudioUtils.hpp> -#include "BaseProcessing.hpp" -#include "Model.hpp" -#include "Classifier.hpp" -#include "MicroNetKwsMfcc.hpp" - -#include <functional> - -namespace arm { -namespace app { - - /** - * @brief Pre-processing class for Keyword Spotting use case. - * Implements methods declared by BasePreProcess and anything else needed - * to populate input tensors ready for inference. - */ - class KwsPreProcess : public BasePreProcess { - - public: - /** - * @brief Constructor - * @param[in] inputTensor Pointer to the TFLite Micro input Tensor. - * @param[in] numFeatures How many MFCC features to use. - * @param[in] numFeatureFrames Number of MFCC vectors that need to be calculated - * for an inference. - * @param[in] mfccFrameLength Number of audio samples used to calculate one set of MFCC values when - * sliding a window through the audio sample. - * @param[in] mfccFrameStride Number of audio samples between consecutive windows. - **/ - explicit KwsPreProcess(TfLiteTensor* inputTensor, size_t numFeatures, size_t numFeatureFrames, - int mfccFrameLength, int mfccFrameStride); - - /** - * @brief Should perform pre-processing of 'raw' input audio data and load it into - * TFLite Micro input tensors ready for inference. - * @param[in] input Pointer to the data that pre-processing will work on. - * @param[in] inputSize Size of the input data. - * @return true if successful, false otherwise. - **/ - bool DoPreProcess(const void* input, size_t inputSize) override; - - size_t m_audioWindowIndex = 0; /* Index of audio slider, used when caching features in longer clips. */ - size_t m_audioDataWindowSize; /* Amount of audio needed for 1 inference. */ - size_t m_audioDataStride; /* Amount of audio to stride across if doing >1 inference in longer clips. */ - - private: - TfLiteTensor* m_inputTensor; /* Model input tensor. */ - const int m_mfccFrameLength; - const int m_mfccFrameStride; - const size_t m_numMfccFrames; /* How many sets of m_numMfccFeats. */ - - audio::MicroNetKwsMFCC m_mfcc; - audio::SlidingWindow<const int16_t> m_mfccSlidingWindow; - size_t m_numMfccVectorsInAudioStride; - size_t m_numReusedMfccVectors; - std::function<void (std::vector<int16_t>&, int, bool, size_t)> m_mfccFeatureCalculator; - - /** - * @brief Returns a function to perform feature calculation and populates input tensor data with - * MFCC data. - * - * Input tensor data type check is performed to choose correct MFCC feature data type. - * If tensor has an integer data type then original features are quantised. - * - * Warning: MFCC calculator provided as input must have the same life scope as returned function. - * - * @param[in] mfcc MFCC feature calculator. - * @param[in,out] inputTensor Input tensor pointer to store calculated features. - * @param[in] cacheSize Size of the feature vectors cache (number of feature vectors). - * @return Function to be called providing audio sample and sliding window index. - */ - std::function<void (std::vector<int16_t>&, int, bool, size_t)> - GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, - TfLiteTensor* inputTensor, - size_t cacheSize); - - template<class T> - std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> - FeatureCalc(TfLiteTensor* inputTensor, size_t cacheSize, - std::function<std::vector<T> (std::vector<int16_t>& )> compute); - }; - - /** - * @brief Post-processing class for Keyword Spotting use case. - * Implements methods declared by BasePostProcess and anything else needed - * to populate result vector. - */ - class KwsPostProcess : public BasePostProcess { - - private: - TfLiteTensor* m_outputTensor; /* Model output tensor. */ - Classifier& m_kwsClassifier; /* KWS Classifier object. */ - const std::vector<std::string>& m_labels; /* KWS Labels. */ - std::vector<ClassificationResult>& m_results; /* Results vector for a single inference. */ - - public: - /** - * @brief Constructor - * @param[in] outputTensor Pointer to the TFLite Micro output Tensor. - * @param[in] classifier Classifier object used to get top N results from classification. - * @param[in] labels Vector of string labels to identify each output of the model. - * @param[in/out] results Vector of classification results to store decoded outputs. - **/ - KwsPostProcess(TfLiteTensor* outputTensor, Classifier& classifier, - const std::vector<std::string>& labels, - std::vector<ClassificationResult>& results); - - /** - * @brief Should perform post-processing of the result of inference then - * populate KWS result data for any later use. - * @return true if successful, false otherwise. - **/ - bool DoPostProcess() override; - }; - -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_PROCESSING_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/include/KwsResult.hpp b/source/use_case/kws_asr/include/KwsResult.hpp deleted file mode 100644 index 45bb790..0000000 --- a/source/use_case/kws_asr/include/KwsResult.hpp +++ /dev/null @@ -1,63 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_RESULT_HPP -#define KWS_RESULT_HPP - -#include "ClassificationResult.hpp" - -#include <vector> - -namespace arm { -namespace app { -namespace kws { - - using ResultVec = std::vector < arm::app::ClassificationResult >; - - /* Structure for holding kws result. */ - class KwsResult { - - public: - ResultVec m_resultVec; /* Container for "thresholded" classification results. */ - float m_timeStamp; /* Audio timestamp for this result. */ - uint32_t m_inferenceNumber; /* Corresponding inference number. */ - float m_threshold; /* Threshold value for `m_resultVec.` */ - - KwsResult() = delete; - KwsResult(ResultVec& resultVec, - const float timestamp, - const uint32_t inferenceIdx, - const float scoreThreshold) { - - this->m_threshold = scoreThreshold; - this->m_timeStamp = timestamp; - this->m_inferenceNumber = inferenceIdx; - - this->m_resultVec = ResultVec(); - for (auto & i : resultVec) { - if (i.m_normalisedVal >= this->m_threshold) { - this->m_resultVec.emplace_back(i); - } - } - } - ~KwsResult() = default; - }; - -} /* namespace kws */ -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_RESULT_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/include/MicroNetKwsMfcc.hpp b/source/use_case/kws_asr/include/MicroNetKwsMfcc.hpp deleted file mode 100644 index af6ba5f..0000000 --- a/source/use_case/kws_asr/include/MicroNetKwsMfcc.hpp +++ /dev/null @@ -1,51 +0,0 @@ -/* - * Copyright (c) 2021-2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_MICRONET_MFCC_HPP -#define KWS_ASR_MICRONET_MFCC_HPP - -#include "Mfcc.hpp" - -namespace arm { -namespace app { -namespace audio { - - /* Class to provide MicroNet specific MFCC calculation requirements. */ - class MicroNetKwsMFCC : public MFCC { - - public: - static constexpr uint32_t ms_defaultSamplingFreq = 16000; - static constexpr uint32_t ms_defaultNumFbankBins = 40; - static constexpr uint32_t ms_defaultMelLoFreq = 20; - static constexpr uint32_t ms_defaultMelHiFreq = 4000; - static constexpr bool ms_defaultUseHtkMethod = true; - - - explicit MicroNetKwsMFCC(const size_t numFeats, const size_t frameLen) - : MFCC(MfccParams( - ms_defaultSamplingFreq, ms_defaultNumFbankBins, - ms_defaultMelLoFreq, ms_defaultMelHiFreq, - numFeats, frameLen, ms_defaultUseHtkMethod)) - {} - MicroNetKwsMFCC() = delete; - ~MicroNetKwsMFCC() = default; - }; - -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_MICRONET_MFCC_HPP */ diff --git a/source/use_case/kws_asr/include/MicroNetKwsModel.hpp b/source/use_case/kws_asr/include/MicroNetKwsModel.hpp deleted file mode 100644 index 22cf916..0000000 --- a/source/use_case/kws_asr/include/MicroNetKwsModel.hpp +++ /dev/null @@ -1,66 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_MICRONETMODEL_HPP -#define KWS_ASR_MICRONETMODEL_HPP - -#include "Model.hpp" - -namespace arm { -namespace app { -namespace kws { - extern const int g_FrameLength; - extern const int g_FrameStride; - extern const float g_ScoreThreshold; - extern const uint32_t g_NumMfcc; - extern const uint32_t g_NumAudioWins; -} /* namespace kws */ -} /* namespace app */ -} /* namespace arm */ - -namespace arm { -namespace app { - class MicroNetKwsModel : public Model { - public: - /* Indices for the expected model - based on input and output tensor shapes */ - static constexpr uint32_t ms_inputRowsIdx = 1; - static constexpr uint32_t ms_inputColsIdx = 2; - static constexpr uint32_t ms_outputRowsIdx = 2; - static constexpr uint32_t ms_outputColsIdx = 3; - - protected: - /** @brief Gets the reference to op resolver interface class. */ - const tflite::MicroOpResolver& GetOpResolver() override; - - /** @brief Adds operations to the op resolver instance. */ - bool EnlistOperations() override; - - const uint8_t* ModelPointer() override; - - size_t ModelSize() override; - - private: - /* Maximum number of individual operations that can be enlisted. */ - static constexpr int ms_maxOpCnt = 7; - - /* A mutable op resolver instance. */ - tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver; - }; - -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_MICRONETMODEL_HPP */ diff --git a/source/use_case/kws_asr/include/OutputDecode.hpp b/source/use_case/kws_asr/include/OutputDecode.hpp deleted file mode 100644 index cea2c33..0000000 --- a/source/use_case/kws_asr/include/OutputDecode.hpp +++ /dev/null @@ -1,40 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_OUTPUT_DECODE_HPP -#define KWS_ASR_OUTPUT_DECODE_HPP - -#include "AsrClassifier.hpp" - -namespace arm { -namespace app { -namespace audio { -namespace asr { - - /** - * @brief Gets the top N classification results from the - * output vector. - * @param[in] vecResults Label output from classifier. - * @return true if successful, false otherwise. - **/ - std::string DecodeOutput(const std::vector<ClassificationResult>& vecResults); - -} /* namespace asr */ -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_OUTPUT_DECODE_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/include/Wav2LetterMfcc.hpp b/source/use_case/kws_asr/include/Wav2LetterMfcc.hpp deleted file mode 100644 index 75d75da..0000000 --- a/source/use_case/kws_asr/include/Wav2LetterMfcc.hpp +++ /dev/null @@ -1,113 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_WAV2LET_MFCC_HPP -#define KWS_ASR_WAV2LET_MFCC_HPP - -#include "Mfcc.hpp" - -namespace arm { -namespace app { -namespace audio { - - /* Class to provide Wav2Letter specific MFCC calculation requirements. */ - class Wav2LetterMFCC : public MFCC { - - public: - static constexpr uint32_t ms_defaultSamplingFreq = 16000; - static constexpr uint32_t ms_defaultNumFbankBins = 128; - static constexpr uint32_t ms_defaultMelLoFreq = 0; - static constexpr uint32_t ms_defaultMelHiFreq = 8000; - static constexpr bool ms_defaultUseHtkMethod = false; - - explicit Wav2LetterMFCC(const size_t numFeats, const size_t frameLen) - : MFCC(MfccParams( - ms_defaultSamplingFreq, ms_defaultNumFbankBins, - ms_defaultMelLoFreq, ms_defaultMelHiFreq, - numFeats, frameLen, ms_defaultUseHtkMethod)) - {} - - Wav2LetterMFCC() = delete; - ~Wav2LetterMFCC() = default; - - protected: - - /** - * @brief Overrides base class implementation of this 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 - * to be used for each bin. - * @param[in] filterBankFilterLast Vector containing the last indices of filter bank - * to be used for each bin. - * @param[out] melEnergies Pre-allocated vector of MEL energies to be - * populated. - * @return true if successful, false otherwise. - */ - bool ApplyMelFilterBank( - std::vector<float>& fftVec, - std::vector<std::vector<float>>& melFilterBank, - std::vector<uint32_t>& filterBankFilterFirst, - std::vector<uint32_t>& filterBankFilterLast, - std::vector<float>& melEnergies) override; - - /** - * @brief Override for the base class implementation convert mel - * energies to logarithmic scale. The difference from - * default behaviour is that the power is converted to dB - * and subsequently clamped. - * @param[in,out] melEnergies 1D vector of Mel energies. - **/ - void ConvertToLogarithmicScale( - std::vector<float>& melEnergies) override; - - /** - * @brief Create a matrix used to calculate Discrete Cosine - * Transform. Override for the base class' default - * implementation as the first and last elements - * use a different normaliser. - * @param[in] inputLength Input length of the buffer on which - * DCT will be performed. - * @param[in] coefficientCount Total coefficients per input length. - * @return 1D vector with inputLength x coefficientCount elements - * populated with DCT coefficients. - */ - std::vector<float> CreateDCTMatrix( - int32_t inputLength, - int32_t coefficientCount) override; - - /** - * @brief Given the low and high Mel values, get the normaliser - * for weights to be applied when populating the filter - * bank. Override for the base class implementation. - * @param[in] leftMel Low Mel frequency value. - * @param[in] rightMel High Mel frequency value. - * @param[in] useHTKMethod Bool to signal if HTK method is to be - * used for calculation. - * @return Value to use for normalising. - */ - float GetMelFilterBankNormaliser( - const float& leftMel, - const float& rightMel, - bool useHTKMethod) override; - - }; - -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_WAV2LET_MFCC_HPP */ diff --git a/source/use_case/kws_asr/include/Wav2LetterModel.hpp b/source/use_case/kws_asr/include/Wav2LetterModel.hpp deleted file mode 100644 index 0e1adc5..0000000 --- a/source/use_case/kws_asr/include/Wav2LetterModel.hpp +++ /dev/null @@ -1,71 +0,0 @@ -/* - * Copyright (c) 2021-2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_WAV2LETTER_MODEL_HPP -#define KWS_ASR_WAV2LETTER_MODEL_HPP - -#include "Model.hpp" - -namespace arm { -namespace app { -namespace asr { - extern const int g_FrameLength; - extern const int g_FrameStride; - extern const float g_ScoreThreshold; - extern const int g_ctxLen; -} /* namespace asr */ -} /* namespace app */ -} /* namespace arm */ - -namespace arm { -namespace app { - - class Wav2LetterModel : public Model { - - public: - /* Indices for the expected model - based on input and output tensor shapes */ - static constexpr uint32_t ms_inputRowsIdx = 1; - static constexpr uint32_t ms_inputColsIdx = 2; - static constexpr uint32_t ms_outputRowsIdx = 2; - static constexpr uint32_t ms_outputColsIdx = 3; - - /* Model specific constants. */ - static constexpr uint32_t ms_blankTokenIdx = 28; - static constexpr uint32_t ms_numMfccFeatures = 13; - - protected: - /** @brief Gets the reference to op resolver interface class. */ - const tflite::MicroOpResolver& GetOpResolver() override; - - /** @brief Adds operations to the op resolver instance. */ - bool EnlistOperations() override; - - const uint8_t* ModelPointer() override; - - size_t ModelSize() override; - - private: - /* Maximum number of individual operations that can be enlisted. */ - static constexpr int ms_maxOpCnt = 5; - - /* A mutable op resolver instance. */ - tflite::MicroMutableOpResolver<ms_maxOpCnt> m_opResolver; - }; - -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_WAV2LETTER_MODEL_HPP */ diff --git a/source/use_case/kws_asr/include/Wav2LetterPostprocess.hpp b/source/use_case/kws_asr/include/Wav2LetterPostprocess.hpp deleted file mode 100644 index d1bc9a2..0000000 --- a/source/use_case/kws_asr/include/Wav2LetterPostprocess.hpp +++ /dev/null @@ -1,108 +0,0 @@ -/* - * Copyright (c) 2021-2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_WAV2LETTER_POSTPROCESS_HPP -#define KWS_ASR_WAV2LETTER_POSTPROCESS_HPP - -#include "TensorFlowLiteMicro.hpp" /* TensorFlow headers. */ -#include "BaseProcessing.hpp" -#include "AsrClassifier.hpp" -#include "AsrResult.hpp" -#include "log_macros.h" - -namespace arm { -namespace app { - - /** - * @brief Helper class to manage tensor post-processing for "wav2letter" - * output. - */ - class AsrPostProcess : public BasePostProcess { - public: - bool m_lastIteration = false; /* Flag to set if processing the last set of data for a clip. */ - - /** - * @brief Constructor - * @param[in] outputTensor Pointer to the TFLite Micro output Tensor. - * @param[in] classifier Object used to get top N results from classification. - * @param[in] labels Vector of string labels to identify each output of the model. - * @param[in/out] result Vector of classification results to store decoded outputs. - * @param[in] outputContextLen Left/right context length for output tensor. - * @param[in] blankTokenIdx Index in the labels that the "Blank token" takes. - * @param[in] reductionAxis The axis that the logits of each time step is on. - **/ - AsrPostProcess(TfLiteTensor* outputTensor, AsrClassifier& classifier, - const std::vector<std::string>& labels, asr::ResultVec& result, - uint32_t outputContextLen, - uint32_t blankTokenIdx, uint32_t reductionAxis); - - /** - * @brief Should perform post-processing of the result of inference then - * populate ASR result data for any later use. - * @return true if successful, false otherwise. - **/ - bool DoPostProcess() override; - - /** @brief Gets the output inner length for post-processing. */ - static uint32_t GetOutputInnerLen(const TfLiteTensor*, uint32_t outputCtxLen); - - /** @brief Gets the output context length (left/right) for post-processing. */ - static uint32_t GetOutputContextLen(const Model& model, uint32_t inputCtxLen); - - /** @brief Gets the number of feature vectors to be computed. */ - static uint32_t GetNumFeatureVectors(const Model& model); - - private: - AsrClassifier& m_classifier; /* ASR Classifier object. */ - TfLiteTensor* m_outputTensor; /* Model output tensor. */ - const std::vector<std::string>& m_labels; /* ASR Labels. */ - asr::ResultVec & m_results; /* Results vector for a single inference. */ - uint32_t m_outputContextLen; /* lengths of left/right contexts for output. */ - uint32_t m_outputInnerLen; /* Length of output inner context. */ - uint32_t m_totalLen; /* Total length of the required axis. */ - uint32_t m_countIterations; /* Current number of iterations. */ - uint32_t m_blankTokenIdx; /* Index of the labels blank token. */ - uint32_t m_reductionAxisIdx; /* Axis containing output logits for a single step. */ - - /** - * @brief Checks if the tensor and axis index are valid - * inputs to the object - based on how it has been initialised. - * @return true if valid, false otherwise. - */ - bool IsInputValid(TfLiteTensor* tensor, - uint32_t axisIdx) const; - - /** - * @brief Gets the tensor data element size in bytes based - * on the tensor type. - * @return Size in bytes, 0 if not supported. - */ - static uint32_t GetTensorElementSize(TfLiteTensor* tensor); - - /** - * @brief Erases sections from the data assuming row-wise - * arrangement along the context axis. - * @return true if successful, false otherwise. - */ - bool EraseSectionsRowWise(uint8_t* ptrData, - uint32_t strideSzBytes, - bool lastIteration); - }; - -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_WAV2LETTER_POSTPROCESS_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/include/Wav2LetterPreprocess.hpp b/source/use_case/kws_asr/include/Wav2LetterPreprocess.hpp deleted file mode 100644 index 1224c23..0000000 --- a/source/use_case/kws_asr/include/Wav2LetterPreprocess.hpp +++ /dev/null @@ -1,182 +0,0 @@ -/* - * Copyright (c) 2021-2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#ifndef KWS_ASR_WAV2LETTER_PREPROCESS_HPP -#define KWS_ASR_WAV2LETTER_PREPROCESS_HPP - -#include "Wav2LetterModel.hpp" -#include "Wav2LetterMfcc.hpp" -#include "AudioUtils.hpp" -#include "DataStructures.hpp" -#include "BaseProcessing.hpp" -#include "log_macros.h" - -namespace arm { -namespace app { - - /* Class to facilitate pre-processing calculation for Wav2Letter model - * for ASR. */ - using AudioWindow = audio::SlidingWindow<const int16_t>; - - class AsrPreProcess : public BasePreProcess { - public: - /** - * @brief Constructor. - * @param[in] inputTensor Pointer to the TFLite Micro input Tensor. - * @param[in] numMfccFeatures Number of MFCC features per window. - * @param[in] numFeatureFrames Number of MFCC vectors that need to be calculated - * for an inference. - * @param[in] mfccWindowLen Number of audio elements to calculate MFCC features per window. - * @param[in] mfccWindowStride Stride (in number of elements) for moving the MFCC window. - */ - AsrPreProcess(TfLiteTensor* inputTensor, - uint32_t numMfccFeatures, - uint32_t numFeatureFrames, - uint32_t mfccWindowLen, - uint32_t mfccWindowStride); - - /** - * @brief Calculates the features required from audio data. This - * includes MFCC, first and second order deltas, - * normalisation and finally, quantisation. The tensor is - * populated with features from a given window placed along - * in a single row. - * @param[in] audioData Pointer to the first element of audio data. - * @param[in] audioDataLen Number of elements in the audio data. - * @return true if successful, false in case of error. - */ - bool DoPreProcess(const void* audioData, size_t audioDataLen) override; - - protected: - /** - * @brief Computes the first and second order deltas for the - * MFCC buffers - they are assumed to be populated. - * - * @param[in] mfcc MFCC buffers. - * @param[out] delta1 Result of the first diff computation. - * @param[out] delta2 Result of the second diff computation. - * @return true if successful, false otherwise. - */ - static bool ComputeDeltas(Array2d<float>& mfcc, - Array2d<float>& delta1, - Array2d<float>& delta2); - - /** - * @brief Given a 2D vector of floats, rescale it to have mean of 0 and - * standard deviation of 1. - * @param[in,out] vec Vector of vector of floats. - */ - static void StandardizeVecF32(Array2d<float>& vec); - - /** - * @brief Standardizes all the MFCC and delta buffers to have mean 0 and std. dev 1. - */ - void Standarize(); - - /** - * @brief Given the quantisation and data type limits, computes - * the quantised values of a floating point input data. - * @param[in] elem Element to be quantised. - * @param[in] quantScale Scale. - * @param[in] quantOffset Offset. - * @param[in] minVal Numerical limit - minimum. - * @param[in] maxVal Numerical limit - maximum. - * @return Floating point quantised value. - */ - static float GetQuantElem( - float elem, - float quantScale, - int quantOffset, - float minVal, - float maxVal); - - /** - * @brief Quantises the MFCC and delta buffers, and places them - * in the output buffer. While doing so, it transposes - * the data. Reason: Buffers in this class are arranged - * for "time" axis to be row major. Primary reason for - * this being the convolution speed up (as we can use - * contiguous memory). The output, however, requires the - * time axis to be in column major arrangement. - * @param[in] outputBuf Pointer to the output buffer. - * @param[in] outputBufSz Output buffer's size. - * @param[in] quantScale Quantisation scale. - * @param[in] quantOffset Quantisation offset. - */ - template <typename T> - bool Quantise( - T* outputBuf, - const uint32_t outputBufSz, - const float quantScale, - const int quantOffset) - { - /* Check the output size will fit everything. */ - if (outputBufSz < (this->m_mfccBuf.size(0) * 3 * sizeof(T))) { - printf_err("Tensor size too small for features\n"); - return false; - } - - /* Populate. */ - T* outputBufMfcc = outputBuf; - T* outputBufD1 = outputBuf + this->m_numMfccFeats; - T* outputBufD2 = outputBufD1 + this->m_numMfccFeats; - const uint32_t ptrIncr = this->m_numMfccFeats * 2; /* (3 vectors - 1 vector) */ - - const float minVal = std::numeric_limits<T>::min(); - const float maxVal = std::numeric_limits<T>::max(); - - /* Need to transpose while copying and concatenating the tensor. */ - for (uint32_t j = 0; j < this->m_numFeatureFrames; ++j) { - for (uint32_t i = 0; i < this->m_numMfccFeats; ++i) { - *outputBufMfcc++ = static_cast<T>(AsrPreProcess::GetQuantElem( - this->m_mfccBuf(i, j), quantScale, - quantOffset, minVal, maxVal)); - *outputBufD1++ = static_cast<T>(AsrPreProcess::GetQuantElem( - this->m_delta1Buf(i, j), quantScale, - quantOffset, minVal, maxVal)); - *outputBufD2++ = static_cast<T>(AsrPreProcess::GetQuantElem( - this->m_delta2Buf(i, j), quantScale, - quantOffset, minVal, maxVal)); - } - outputBufMfcc += ptrIncr; - outputBufD1 += ptrIncr; - outputBufD2 += ptrIncr; - } - - return true; - } - - private: - audio::Wav2LetterMFCC m_mfcc; /* MFCC instance. */ - TfLiteTensor* m_inputTensor; /* Model input tensor. */ - - /* Actual buffers to be populated. */ - Array2d<float> m_mfccBuf; /* Contiguous buffer 1D: MFCC */ - Array2d<float> m_delta1Buf; /* Contiguous buffer 1D: Delta 1 */ - Array2d<float> m_delta2Buf; /* Contiguous buffer 1D: Delta 2 */ - - uint32_t m_mfccWindowLen; /* Window length for MFCC. */ - uint32_t m_mfccWindowStride; /* Window stride len for MFCC. */ - uint32_t m_numMfccFeats; /* Number of MFCC features per window. */ - uint32_t m_numFeatureFrames; /* How many sets of m_numMfccFeats. */ - AudioWindow m_mfccSlidingWindow; /* Sliding window to calculate MFCCs. */ - - }; - -} /* namespace app */ -} /* namespace arm */ - -#endif /* KWS_ASR_WAV2LETTER_PREPROCESS_HPP */
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/AsrClassifier.cc b/source/use_case/kws_asr/src/AsrClassifier.cc deleted file mode 100644 index 9c18b14..0000000 --- a/source/use_case/kws_asr/src/AsrClassifier.cc +++ /dev/null @@ -1,136 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "AsrClassifier.hpp" - -#include "log_macros.h" -#include "TensorFlowLiteMicro.hpp" -#include "Wav2LetterModel.hpp" - -template<typename T> -bool arm::app::AsrClassifier::GetTopResults(TfLiteTensor* tensor, - std::vector<ClassificationResult>& vecResults, - const std::vector <std::string>& labels, double scale, double zeroPoint) -{ - const uint32_t nElems = tensor->dims->data[arm::app::Wav2LetterModel::ms_outputRowsIdx]; - const uint32_t nLetters = tensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx]; - - if (nLetters != labels.size()) { - printf("Output size doesn't match the labels' size\n"); - return false; - } - - /* NOTE: tensor's size verification against labels should be - * checked by the calling/public function. */ - if (nLetters < 1) { - return false; - } - - /* Final results' container. */ - vecResults = std::vector<ClassificationResult>(nElems); - - T* tensorData = tflite::GetTensorData<T>(tensor); - - /* Get the top 1 results. */ - for (uint32_t i = 0, row = 0; i < nElems; ++i, row+=nLetters) { - std::pair<T, uint32_t> top_1 = std::make_pair(tensorData[row], 0); - - for (uint32_t j = 1; j < nLetters; ++j) { - if (top_1.first < tensorData[row + j]) { - top_1.first = tensorData[row + j]; - top_1.second = j; - } - } - - double score = static_cast<int> (top_1.first); - vecResults[i].m_normalisedVal = scale * (score - zeroPoint); - vecResults[i].m_label = labels[top_1.second]; - vecResults[i].m_labelIdx = top_1.second; - } - - return true; -} -template bool arm::app::AsrClassifier::GetTopResults<uint8_t>(TfLiteTensor* tensor, - std::vector<ClassificationResult>& vecResults, - const std::vector <std::string>& labels, double scale, double zeroPoint); -template bool arm::app::AsrClassifier::GetTopResults<int8_t>(TfLiteTensor* tensor, - std::vector<ClassificationResult>& vecResults, - const std::vector <std::string>& labels, double scale, double zeroPoint); - -bool arm::app::AsrClassifier::GetClassificationResults( - TfLiteTensor* outputTensor, - std::vector<ClassificationResult>& vecResults, - const std::vector <std::string>& labels, uint32_t topNCount, bool use_softmax) -{ - UNUSED(use_softmax); - vecResults.clear(); - - constexpr int minTensorDims = static_cast<int>( - (arm::app::Wav2LetterModel::ms_outputRowsIdx > arm::app::Wav2LetterModel::ms_outputColsIdx)? - arm::app::Wav2LetterModel::ms_outputRowsIdx : arm::app::Wav2LetterModel::ms_outputColsIdx); - - constexpr uint32_t outColsIdx = arm::app::Wav2LetterModel::ms_outputColsIdx; - - /* Sanity checks. */ - if (outputTensor == nullptr) { - printf_err("Output vector is null pointer.\n"); - return false; - } else if (outputTensor->dims->size < minTensorDims) { - printf_err("Output tensor expected to be 3D (1, m, n)\n"); - return false; - } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) < topNCount) { - printf_err("Output vectors are smaller than %" PRIu32 "\n", topNCount); - return false; - } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) != labels.size()) { - printf("Output size doesn't match the labels' size\n"); - return false; - } - - if (topNCount != 1) { - warn("TopNCount value ignored in this implementation\n"); - } - - /* To return the floating point values, we need quantization parameters. */ - QuantParams quantParams = GetTensorQuantParams(outputTensor); - - bool resultState; - - switch (outputTensor->type) { - case kTfLiteUInt8: - resultState = this->GetTopResults<uint8_t>( - outputTensor, vecResults, - labels, quantParams.scale, - quantParams.offset); - break; - case kTfLiteInt8: - resultState = this->GetTopResults<int8_t>( - outputTensor, vecResults, - labels, quantParams.scale, - quantParams.offset); - break; - default: - printf_err("Tensor type %s not supported by classifier\n", - TfLiteTypeGetName(outputTensor->type)); - return false; - } - - if (!resultState) { - printf_err("Failed to get sorted set\n"); - return false; - } - - return true; -}
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/KwsProcessing.cc b/source/use_case/kws_asr/src/KwsProcessing.cc deleted file mode 100644 index 328709d..0000000 --- a/source/use_case/kws_asr/src/KwsProcessing.cc +++ /dev/null @@ -1,212 +0,0 @@ -/* - * Copyright (c) 2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "KwsProcessing.hpp" -#include "ImageUtils.hpp" -#include "log_macros.h" -#include "MicroNetKwsModel.hpp" - -namespace arm { -namespace app { - - KwsPreProcess::KwsPreProcess(TfLiteTensor* inputTensor, size_t numFeatures, size_t numMfccFrames, - int mfccFrameLength, int mfccFrameStride - ): - m_inputTensor{inputTensor}, - m_mfccFrameLength{mfccFrameLength}, - m_mfccFrameStride{mfccFrameStride}, - m_numMfccFrames{numMfccFrames}, - m_mfcc{audio::MicroNetKwsMFCC(numFeatures, mfccFrameLength)} - { - this->m_mfcc.Init(); - - /* Deduce the data length required for 1 inference from the network parameters. */ - this->m_audioDataWindowSize = this->m_numMfccFrames * this->m_mfccFrameStride + - (this->m_mfccFrameLength - this->m_mfccFrameStride); - - /* Creating an MFCC feature sliding window for the data required for 1 inference. */ - this->m_mfccSlidingWindow = audio::SlidingWindow<const int16_t>(nullptr, this->m_audioDataWindowSize, - this->m_mfccFrameLength, this->m_mfccFrameStride); - - /* For longer audio clips we choose to move by half the audio window size - * => for a 1 second window size there is an overlap of 0.5 seconds. */ - this->m_audioDataStride = this->m_audioDataWindowSize / 2; - - /* To have the previously calculated features re-usable, stride must be multiple - * of MFCC features window stride. Reduce stride through audio if needed. */ - if (0 != this->m_audioDataStride % this->m_mfccFrameStride) { - this->m_audioDataStride -= this->m_audioDataStride % this->m_mfccFrameStride; - } - - this->m_numMfccVectorsInAudioStride = this->m_audioDataStride / this->m_mfccFrameStride; - - /* Calculate number of the feature vectors in the window overlap region. - * These feature vectors will be reused.*/ - this->m_numReusedMfccVectors = this->m_mfccSlidingWindow.TotalStrides() + 1 - - this->m_numMfccVectorsInAudioStride; - - /* Construct feature calculation function. */ - this->m_mfccFeatureCalculator = GetFeatureCalculator(this->m_mfcc, this->m_inputTensor, - this->m_numReusedMfccVectors); - - if (!this->m_mfccFeatureCalculator) { - printf_err("Feature calculator not initialized."); - } - } - - bool KwsPreProcess::DoPreProcess(const void* data, size_t inputSize) - { - UNUSED(inputSize); - if (data == nullptr) { - printf_err("Data pointer is null"); - } - - /* Set the features sliding window to the new address. */ - auto input = static_cast<const int16_t*>(data); - this->m_mfccSlidingWindow.Reset(input); - - /* Cache is only usable if we have more than 1 inference in an audio clip. */ - bool useCache = this->m_audioWindowIndex > 0 && this->m_numReusedMfccVectors > 0; - - /* Use a sliding window to calculate MFCC features frame by frame. */ - while (this->m_mfccSlidingWindow.HasNext()) { - const int16_t* mfccWindow = this->m_mfccSlidingWindow.Next(); - - std::vector<int16_t> mfccFrameAudioData = std::vector<int16_t>(mfccWindow, - mfccWindow + this->m_mfccFrameLength); - - /* Compute features for this window and write them to input tensor. */ - this->m_mfccFeatureCalculator(mfccFrameAudioData, this->m_mfccSlidingWindow.Index(), - useCache, this->m_numMfccVectorsInAudioStride); - } - - debug("Input tensor populated \n"); - - return true; - } - - /** - * @brief Generic feature calculator factory. - * - * Returns lambda function to compute features using features cache. - * Real features math is done by a lambda function provided as a parameter. - * Features are written to input tensor memory. - * - * @tparam T Feature vector type. - * @param[in] inputTensor Model input tensor pointer. - * @param[in] cacheSize Number of feature vectors to cache. Defined by the sliding window overlap. - * @param[in] compute Features calculator function. - * @return Lambda function to compute features. - */ - template<class T> - std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> - KwsPreProcess::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); - - return [=](std::vector<int16_t>& audioDataWindow, - size_t index, - bool useCache, - size_t featuresOverlapIndex) - { - T* tensorData = tflite::GetTensorData<T>(inputTensor); - std::vector<T> features; - - /* Reuse features from cache if cache is ready and sliding windows overlap. - * Overlap is in the beginning of sliding window with a size of a feature cache. */ - if (useCache && index < featureCache.size()) { - features = std::move(featureCache[index]); - } else { - features = std::move(compute(audioDataWindow)); - } - auto size = features.size(); - auto sizeBytes = sizeof(T) * size; - std::memcpy(tensorData + (index * size), features.data(), sizeBytes); - - /* Start renewing cache as soon iteration goes out of the windows overlap. */ - if (index >= featuresOverlapIndex) { - featureCache[index - featuresOverlapIndex] = std::move(features); - } - }; - } - - template std::function<void (std::vector<int16_t>&, size_t , bool, size_t)> - KwsPreProcess::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)> - KwsPreProcess::FeatureCalc<float>(TfLiteTensor* inputTensor, - size_t cacheSize, - std::function<std::vector<float>(std::vector<int16_t>&)> compute); - - - std::function<void (std::vector<int16_t>&, int, bool, size_t)> - KwsPreProcess::GetFeatureCalculator(audio::MicroNetKwsMFCC& mfcc, TfLiteTensor* inputTensor, size_t cacheSize) - { - std::function<void (std::vector<int16_t>&, size_t, bool, size_t)> mfccFeatureCalc; - - TfLiteQuantization quant = inputTensor->quantization; - - if (kTfLiteAffineQuantization == quant.type) { - auto *quantParams = (TfLiteAffineQuantization *) quant.params; - const float quantScale = quantParams->scale->data[0]; - const int quantOffset = quantParams->zero_point->data[0]; - - switch (inputTensor->type) { - case kTfLiteInt8: { - mfccFeatureCalc = this->FeatureCalc<int8_t>(inputTensor, - cacheSize, - [=, &mfcc](std::vector<int16_t>& audioDataWindow) { - return mfcc.MfccComputeQuant<int8_t>(audioDataWindow, - quantScale, - quantOffset); - } - ); - break; - } - default: - printf_err("Tensor type %s not supported\n", TfLiteTypeGetName(inputTensor->type)); - } - } else { - mfccFeatureCalc = this->FeatureCalc<float>(inputTensor, cacheSize, - [&mfcc](std::vector<int16_t>& audioDataWindow) { - return mfcc.MfccCompute(audioDataWindow); } - ); - } - return mfccFeatureCalc; - } - - KwsPostProcess::KwsPostProcess(TfLiteTensor* outputTensor, Classifier& classifier, - const std::vector<std::string>& labels, - std::vector<ClassificationResult>& results) - :m_outputTensor{outputTensor}, - m_kwsClassifier{classifier}, - m_labels{labels}, - m_results{results} - {} - - bool KwsPostProcess::DoPostProcess() - { - return this->m_kwsClassifier.GetClassificationResults( - this->m_outputTensor, this->m_results, - this->m_labels, 1, true); - } - -} /* namespace app */ -} /* namespace arm */
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/MainLoop.cc b/source/use_case/kws_asr/src/MainLoop.cc index f1d97a0..2365264 100644 --- a/source/use_case/kws_asr/src/MainLoop.cc +++ b/source/use_case/kws_asr/src/MainLoop.cc @@ -23,7 +23,24 @@ #include "Wav2LetterModel.hpp" /* ASR model class for running inference. */ #include "UseCaseCommonUtils.hpp" /* Utils functions. */ #include "UseCaseHandler.hpp" /* Handlers for different user options. */ -#include "log_macros.h" +#include "log_macros.h" /* Logging functions */ +#include "BufAttributes.hpp" /* Buffer attributes to be applied */ + +namespace arm { +namespace app { + static uint8_t tensorArena[ACTIVATION_BUF_SZ] ACTIVATION_BUF_ATTRIBUTE; + + namespace asr { + extern uint8_t* GetModelPointer(); + extern size_t GetModelLen(); + } + + namespace kws { + extern uint8_t* GetModelPointer(); + extern size_t GetModelLen(); + } +} /* namespace app */ +} /* namespace arm */ using KwsClassifier = arm::app::Classifier; @@ -60,14 +77,29 @@ void main_loop() arm::app::Wav2LetterModel asrModel; /* Load the models. */ - if (!kwsModel.Init()) { + if (!kwsModel.Init(arm::app::tensorArena, + sizeof(arm::app::tensorArena), + arm::app::kws::GetModelPointer(), + arm::app::kws::GetModelLen())) { printf_err("Failed to initialise KWS model\n"); return; } +#if !defined(ARM_NPU) + /* If it is not a NPU build check if the model contains a NPU operator */ + if (kwsModel.ContainsEthosUOperator()) { + printf_err("No driver support for Ethos-U operator found in the KWS model.\n"); + return; + } +#endif /* ARM_NPU */ + /* Initialise the asr model using the same allocator from KWS * to re-use the tensor arena. */ - if (!asrModel.Init(kwsModel.GetAllocator())) { + if (!asrModel.Init(arm::app::tensorArena, + sizeof(arm::app::tensorArena), + arm::app::asr::GetModelPointer(), + arm::app::asr::GetModelLen(), + kwsModel.GetAllocator())) { printf_err("Failed to initialise ASR model\n"); return; } else if (!VerifyTensorDimensions(asrModel)) { @@ -75,6 +107,14 @@ void main_loop() return; } +#if !defined(ARM_NPU) + /* If it is not a NPU build check if the model contains a NPU operator */ + if (asrModel.ContainsEthosUOperator()) { + printf_err("No driver support for Ethos-U operator found in the ASR model.\n"); + return; + } +#endif /* ARM_NPU */ + /* Instantiate application context. */ arm::app::ApplicationContext caseContext; diff --git a/source/use_case/kws_asr/src/MicroNetKwsModel.cc b/source/use_case/kws_asr/src/MicroNetKwsModel.cc deleted file mode 100644 index 663faa0..0000000 --- a/source/use_case/kws_asr/src/MicroNetKwsModel.cc +++ /dev/null @@ -1,63 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "MicroNetKwsModel.hpp" -#include "log_macros.h" - -namespace arm { -namespace app { -namespace kws { - extern uint8_t* GetModelPointer(); - extern size_t GetModelLen(); -} /* namespace kws */ -} /* namespace app */ -} /* namespace arm */ - -const tflite::MicroOpResolver& arm::app::MicroNetKwsModel::GetOpResolver() -{ - return this->m_opResolver; -} - -bool arm::app::MicroNetKwsModel::EnlistOperations() -{ - this->m_opResolver.AddAveragePool2D(); - this->m_opResolver.AddConv2D(); - this->m_opResolver.AddDepthwiseConv2D(); - this->m_opResolver.AddFullyConnected(); - this->m_opResolver.AddRelu(); - this->m_opResolver.AddReshape(); - -#if defined(ARM_NPU) - if (kTfLiteOk == this->m_opResolver.AddEthosU()) { - info("Added %s support to op resolver\n", - tflite::GetString_ETHOSU()); - } else { - printf_err("Failed to add Arm NPU support to op resolver."); - return false; - } -#endif /* ARM_NPU */ - return true; -} - -const uint8_t* arm::app::MicroNetKwsModel::ModelPointer() -{ - return arm::app::kws::GetModelPointer(); -} - -size_t arm::app::MicroNetKwsModel::ModelSize() -{ - return arm::app::kws::GetModelLen(); -}
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/OutputDecode.cc b/source/use_case/kws_asr/src/OutputDecode.cc deleted file mode 100644 index 41fbe07..0000000 --- a/source/use_case/kws_asr/src/OutputDecode.cc +++ /dev/null @@ -1,47 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "OutputDecode.hpp" - -namespace arm { -namespace app { -namespace audio { -namespace asr { - - std::string DecodeOutput(const std::vector<ClassificationResult>& vecResults) - { - std::string CleanOutputBuffer; - - for (size_t i = 0; i < vecResults.size(); ++i) /* For all elements in vector. */ - { - while (i+1 < vecResults.size() && - vecResults[i].m_label == vecResults[i+1].m_label) /* While the current element is equal to the next, ignore it and move on. */ - { - ++i; - } - if (vecResults[i].m_label != "$") /* $ is a character used to represent unknown and double characters so should not be in output. */ - { - CleanOutputBuffer += vecResults[i].m_label; /* If the element is different to the next, it will be appended to CleanOutputBuffer. */ - } - } - - return CleanOutputBuffer; /* Return string type containing clean output. */ - } - -} /* namespace asr */ -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ diff --git a/source/use_case/kws_asr/src/UseCaseHandler.cc b/source/use_case/kws_asr/src/UseCaseHandler.cc index 01aefae..9427ae0 100644 --- a/source/use_case/kws_asr/src/UseCaseHandler.cc +++ b/source/use_case/kws_asr/src/UseCaseHandler.cc @@ -25,6 +25,7 @@ #include "MicroNetKwsMfcc.hpp" #include "Classifier.hpp" #include "KwsResult.hpp" +#include "Wav2LetterModel.hpp" #include "Wav2LetterMfcc.hpp" #include "Wav2LetterPreprocess.hpp" #include "Wav2LetterPostprocess.hpp" @@ -470,4 +471,4 @@ namespace app { } } /* namespace app */ -} /* namespace arm */
\ No newline at end of file +} /* namespace arm */ diff --git a/source/use_case/kws_asr/src/Wav2LetterMfcc.cc b/source/use_case/kws_asr/src/Wav2LetterMfcc.cc deleted file mode 100644 index f2c50f3..0000000 --- a/source/use_case/kws_asr/src/Wav2LetterMfcc.cc +++ /dev/null @@ -1,141 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "Wav2LetterMfcc.hpp" - -#include "PlatformMath.hpp" -#include "log_macros.h" - -#include <cfloat> - -namespace arm { -namespace app { -namespace audio { - - bool Wav2LetterMFCC::ApplyMelFilterBank( - std::vector<float>& fftVec, - std::vector<std::vector<float>>& melFilterBank, - std::vector<uint32_t>& filterBankFilterFirst, - std::vector<uint32_t>& filterBankFilterLast, - std::vector<float>& melEnergies) - { - const size_t numBanks = melEnergies.size(); - - if (numBanks != filterBankFilterFirst.size() || - numBanks != filterBankFilterLast.size()) { - printf_err("unexpected filter bank lengths\n"); - return false; - } - - for (size_t bin = 0; bin < numBanks; ++bin) { - auto filterBankIter = melFilterBank[bin].begin(); - auto end = melFilterBank[bin].end(); - /* Avoid log of zero at later stages, same value used in librosa. - * The number was used during our default wav2letter model training. */ - float melEnergy = 1e-10; - const uint32_t firstIndex = filterBankFilterFirst[bin]; - const uint32_t lastIndex = std::min<uint32_t>(filterBankFilterLast[bin], fftVec.size() - 1); - - for (uint32_t i = firstIndex; i <= lastIndex && filterBankIter != end; ++i) { - melEnergy += (*filterBankIter++ * fftVec[i]); - } - - melEnergies[bin] = melEnergy; - } - - return true; - } - - void Wav2LetterMFCC::ConvertToLogarithmicScale( - std::vector<float>& melEnergies) - { - float maxMelEnergy = -FLT_MAX; - - /* Container for natural logarithms of mel energies. */ - std::vector <float> vecLogEnergies(melEnergies.size(), 0.f); - - /* Because we are taking natural logs, we need to multiply by log10(e). - * Also, for wav2letter model, we scale our log10 values by 10. */ - constexpr float multiplier = 10.0 * /* Default scalar. */ - 0.4342944819032518; /* log10f(std::exp(1.0))*/ - - /* Take log of the whole vector. */ - math::MathUtils::VecLogarithmF32(melEnergies, vecLogEnergies); - - /* Scale the log values and get the max. */ - for (auto iterM = melEnergies.begin(), iterL = vecLogEnergies.begin(); - iterM != melEnergies.end() && iterL != vecLogEnergies.end(); ++iterM, ++iterL) { - - *iterM = *iterL * multiplier; - - /* Save the max mel energy. */ - if (*iterM > maxMelEnergy) { - maxMelEnergy = *iterM; - } - } - - /* Clamp the mel energies. */ - constexpr float maxDb = 80.0; - const float clampLevelLowdB = maxMelEnergy - maxDb; - for (float & melEnergie : melEnergies) { - melEnergie = std::max(melEnergie, clampLevelLowdB); - } - } - - std::vector<float> Wav2LetterMFCC::CreateDCTMatrix( - const int32_t inputLength, - const int32_t coefficientCount) - { - std::vector<float> dctMatix(inputLength * coefficientCount); - - /* Orthonormal normalization. */ - const float normalizerK0 = 2 * math::MathUtils::SqrtF32(1.0f / - static_cast<float>(4*inputLength)); - const float normalizer = 2 * math::MathUtils::SqrtF32(1.0f / - static_cast<float>(2*inputLength)); - - const float angleIncr = M_PI/inputLength; - float angle = angleIncr; /* We start using it at k = 1 loop. */ - - /* First row of DCT will use normalizer K0 */ - for (int32_t n = 0; n < inputLength; ++n) { - dctMatix[n] = normalizerK0 /* cos(0) = 1 */; - } - - /* Second row (index = 1) onwards, we use standard normalizer. */ - for (int32_t k = 1, m = inputLength; k < coefficientCount; ++k, m += inputLength) { - for (int32_t n = 0; n < inputLength; ++n) { - dctMatix[m+n] = normalizer * - math::MathUtils::CosineF32((n + 0.5f) * angle); - } - angle += angleIncr; - } - return dctMatix; - } - - float Wav2LetterMFCC::GetMelFilterBankNormaliser( - const float& leftMel, - const float& rightMel, - const bool useHTKMethod) - { - /* Slaney normalization for mel weights. */ - return (2.0f / (MFCC::InverseMelScale(rightMel, useHTKMethod) - - MFCC::InverseMelScale(leftMel, useHTKMethod))); - } - -} /* namespace audio */ -} /* namespace app */ -} /* namespace arm */ diff --git a/source/use_case/kws_asr/src/Wav2LetterModel.cc b/source/use_case/kws_asr/src/Wav2LetterModel.cc deleted file mode 100644 index 52bd23a..0000000 --- a/source/use_case/kws_asr/src/Wav2LetterModel.cc +++ /dev/null @@ -1,61 +0,0 @@ -/* - * Copyright (c) 2021 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "Wav2LetterModel.hpp" -#include "log_macros.h" - -namespace arm { -namespace app { -namespace asr { - extern uint8_t* GetModelPointer(); - extern size_t GetModelLen(); -} -} /* namespace app */ -} /* namespace arm */ - -const tflite::MicroOpResolver& arm::app::Wav2LetterModel::GetOpResolver() -{ - return this->m_opResolver; -} - -bool arm::app::Wav2LetterModel::EnlistOperations() -{ - this->m_opResolver.AddConv2D(); - this->m_opResolver.AddLeakyRelu(); - this->m_opResolver.AddSoftmax(); - this->m_opResolver.AddReshape(); - -#if defined(ARM_NPU) - if (kTfLiteOk == this->m_opResolver.AddEthosU()) { - info("Added %s support to op resolver\n", - tflite::GetString_ETHOSU()); - } else { - printf_err("Failed to add Arm NPU support to op resolver."); - return false; - } -#endif /* ARM_NPU */ - return true; -} - -const uint8_t* arm::app::Wav2LetterModel::ModelPointer() -{ - return arm::app::asr::GetModelPointer(); -} - -size_t arm::app::Wav2LetterModel::ModelSize() -{ - return arm::app::asr::GetModelLen(); -}
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/Wav2LetterPostprocess.cc b/source/use_case/kws_asr/src/Wav2LetterPostprocess.cc deleted file mode 100644 index 42f434e..0000000 --- a/source/use_case/kws_asr/src/Wav2LetterPostprocess.cc +++ /dev/null @@ -1,214 +0,0 @@ -/* - * Copyright (c) 2021-2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "Wav2LetterPostprocess.hpp" - -#include "Wav2LetterModel.hpp" -#include "log_macros.h" - -#include <cmath> - -namespace arm { -namespace app { - - AsrPostProcess::AsrPostProcess(TfLiteTensor* outputTensor, AsrClassifier& classifier, - const std::vector<std::string>& labels, std::vector<ClassificationResult>& results, - const uint32_t outputContextLen, - const uint32_t blankTokenIdx, const uint32_t reductionAxisIdx - ): - m_classifier(classifier), - m_outputTensor(outputTensor), - m_labels{labels}, - m_results(results), - m_outputContextLen(outputContextLen), - m_countIterations(0), - m_blankTokenIdx(blankTokenIdx), - m_reductionAxisIdx(reductionAxisIdx) - { - this->m_outputInnerLen = AsrPostProcess::GetOutputInnerLen(this->m_outputTensor, this->m_outputContextLen); - this->m_totalLen = (2 * this->m_outputContextLen + this->m_outputInnerLen); - } - - bool AsrPostProcess::DoPostProcess() - { - /* Basic checks. */ - if (!this->IsInputValid(this->m_outputTensor, this->m_reductionAxisIdx)) { - return false; - } - - /* Irrespective of tensor type, we use unsigned "byte" */ - auto* ptrData = tflite::GetTensorData<uint8_t>(this->m_outputTensor); - const uint32_t elemSz = AsrPostProcess::GetTensorElementSize(this->m_outputTensor); - - /* Other sanity checks. */ - if (0 == elemSz) { - printf_err("Tensor type not supported for post processing\n"); - return false; - } else if (elemSz * this->m_totalLen > this->m_outputTensor->bytes) { - printf_err("Insufficient number of tensor bytes\n"); - return false; - } - - /* Which axis do we need to process? */ - switch (this->m_reductionAxisIdx) { - case Wav2LetterModel::ms_outputRowsIdx: - this->EraseSectionsRowWise( - ptrData, elemSz * this->m_outputTensor->dims->data[Wav2LetterModel::ms_outputColsIdx], - this->m_lastIteration); - break; - default: - printf_err("Unsupported axis index: %" PRIu32 "\n", this->m_reductionAxisIdx); - return false; - } - this->m_classifier.GetClassificationResults(this->m_outputTensor, - this->m_results, this->m_labels, 1); - - return true; - } - - bool AsrPostProcess::IsInputValid(TfLiteTensor* tensor, const uint32_t axisIdx) const - { - if (nullptr == tensor) { - return false; - } - - if (static_cast<int>(axisIdx) >= tensor->dims->size) { - printf_err("Invalid axis index: %" PRIu32 "; Max: %d\n", - axisIdx, tensor->dims->size); - return false; - } - - if (static_cast<int>(this->m_totalLen) != - tensor->dims->data[axisIdx]) { - printf_err("Unexpected tensor dimension for axis %d, got %d, \n", - axisIdx, tensor->dims->data[axisIdx]); - return false; - } - - return true; - } - - uint32_t AsrPostProcess::GetTensorElementSize(TfLiteTensor* tensor) - { - switch(tensor->type) { - case kTfLiteUInt8: - case kTfLiteInt8: - return 1; - case kTfLiteInt16: - return 2; - case kTfLiteInt32: - case kTfLiteFloat32: - return 4; - default: - printf_err("Unsupported tensor type %s\n", - TfLiteTypeGetName(tensor->type)); - } - - return 0; - } - - bool AsrPostProcess::EraseSectionsRowWise( - uint8_t* ptrData, - const uint32_t strideSzBytes, - const bool lastIteration) - { - /* In this case, the "zero-ing" is quite simple as the region - * to be zeroed sits in contiguous memory (row-major). */ - const uint32_t eraseLen = strideSzBytes * this->m_outputContextLen; - - /* Erase left context? */ - if (this->m_countIterations > 0) { - /* Set output of each classification window to the blank token. */ - std::memset(ptrData, 0, eraseLen); - for (size_t windowIdx = 0; windowIdx < this->m_outputContextLen; windowIdx++) { - ptrData[windowIdx*strideSzBytes + this->m_blankTokenIdx] = 1; - } - } - - /* Erase right context? */ - if (false == lastIteration) { - uint8_t* rightCtxPtr = ptrData + (strideSzBytes * (this->m_outputContextLen + this->m_outputInnerLen)); - /* Set output of each classification window to the blank token. */ - std::memset(rightCtxPtr, 0, eraseLen); - for (size_t windowIdx = 0; windowIdx < this->m_outputContextLen; windowIdx++) { - rightCtxPtr[windowIdx*strideSzBytes + this->m_blankTokenIdx] = 1; - } - } - - if (lastIteration) { - this->m_countIterations = 0; - } else { - ++this->m_countIterations; - } - - return true; - } - - uint32_t AsrPostProcess::GetNumFeatureVectors(const Model& model) - { - TfLiteTensor* inputTensor = model.GetInputTensor(0); - const int inputRows = std::max(inputTensor->dims->data[Wav2LetterModel::ms_inputRowsIdx], 0); - if (inputRows == 0) { - printf_err("Error getting number of input rows for axis: %" PRIu32 "\n", - Wav2LetterModel::ms_inputRowsIdx); - } - return inputRows; - } - - uint32_t AsrPostProcess::GetOutputInnerLen(const TfLiteTensor* outputTensor, const uint32_t outputCtxLen) - { - const uint32_t outputRows = std::max(outputTensor->dims->data[Wav2LetterModel::ms_outputRowsIdx], 0); - if (outputRows == 0) { - printf_err("Error getting number of output rows for axis: %" PRIu32 "\n", - Wav2LetterModel::ms_outputRowsIdx); - } - - /* Watching for underflow. */ - int innerLen = (outputRows - (2 * outputCtxLen)); - - return std::max(innerLen, 0); - } - - uint32_t AsrPostProcess::GetOutputContextLen(const Model& model, const uint32_t inputCtxLen) - { - const uint32_t inputRows = AsrPostProcess::GetNumFeatureVectors(model); - const uint32_t inputInnerLen = inputRows - (2 * inputCtxLen); - constexpr uint32_t ms_outputRowsIdx = Wav2LetterModel::ms_outputRowsIdx; - - /* Check to make sure that the input tensor supports the above - * context and inner lengths. */ - if (inputRows <= 2 * inputCtxLen || inputRows <= inputInnerLen) { - printf_err("Input rows not compatible with ctx of %" PRIu32 "\n", - inputCtxLen); - return 0; - } - - TfLiteTensor* outputTensor = model.GetOutputTensor(0); - const uint32_t outputRows = std::max(outputTensor->dims->data[ms_outputRowsIdx], 0); - if (outputRows == 0) { - printf_err("Error getting number of output rows for axis: %" PRIu32 "\n", - Wav2LetterModel::ms_outputRowsIdx); - return 0; - } - - const float inOutRowRatio = static_cast<float>(inputRows) / - static_cast<float>(outputRows); - - return std::round(static_cast<float>(inputCtxLen) / inOutRowRatio); - } - -} /* namespace app */ -} /* namespace arm */
\ No newline at end of file diff --git a/source/use_case/kws_asr/src/Wav2LetterPreprocess.cc b/source/use_case/kws_asr/src/Wav2LetterPreprocess.cc deleted file mode 100644 index 92b0631..0000000 --- a/source/use_case/kws_asr/src/Wav2LetterPreprocess.cc +++ /dev/null @@ -1,208 +0,0 @@ -/* - * Copyright (c) 2021-2022 Arm Limited. All rights reserved. - * SPDX-License-Identifier: Apache-2.0 - * - * Licensed under the Apache License, Version 2.0 (the "License"); - * you may not use this file except in compliance with the License. - * You may obtain a copy of the License at - * - * http://www.apache.org/licenses/LICENSE-2.0 - * - * Unless required by applicable law or agreed to in writing, software - * distributed under the License is distributed on an "AS IS" BASIS, - * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. - * See the License for the specific language governing permissions and - * limitations under the License. - */ -#include "Wav2LetterPreprocess.hpp" - -#include "PlatformMath.hpp" -#include "TensorFlowLiteMicro.hpp" - -#include <algorithm> -#include <cmath> - -namespace arm { -namespace app { - - AsrPreProcess::AsrPreProcess(TfLiteTensor* inputTensor, const uint32_t numMfccFeatures, - const uint32_t numFeatureFrames, const uint32_t mfccWindowLen, - const uint32_t mfccWindowStride - ): - m_mfcc(numMfccFeatures, mfccWindowLen), - m_inputTensor(inputTensor), - m_mfccBuf(numMfccFeatures, numFeatureFrames), - m_delta1Buf(numMfccFeatures, numFeatureFrames), - m_delta2Buf(numMfccFeatures, numFeatureFrames), - m_mfccWindowLen(mfccWindowLen), - m_mfccWindowStride(mfccWindowStride), - m_numMfccFeats(numMfccFeatures), - m_numFeatureFrames(numFeatureFrames) - { - if (numMfccFeatures > 0 && mfccWindowLen > 0) { - this->m_mfcc.Init(); - } - } - - bool AsrPreProcess::DoPreProcess(const void* audioData, const size_t audioDataLen) - { - this->m_mfccSlidingWindow = audio::SlidingWindow<const int16_t>( - static_cast<const int16_t*>(audioData), audioDataLen, - this->m_mfccWindowLen, this->m_mfccWindowStride); - - uint32_t mfccBufIdx = 0; - - std::fill(m_mfccBuf.begin(), m_mfccBuf.end(), 0.f); - std::fill(m_delta1Buf.begin(), m_delta1Buf.end(), 0.f); - std::fill(m_delta2Buf.begin(), m_delta2Buf.end(), 0.f); - - /* While we can slide over the audio. */ - while (this->m_mfccSlidingWindow.HasNext()) { - const int16_t* mfccWindow = this->m_mfccSlidingWindow.Next(); - auto mfccAudioData = std::vector<int16_t>( - mfccWindow, - mfccWindow + this->m_mfccWindowLen); - auto mfcc = this->m_mfcc.MfccCompute(mfccAudioData); - for (size_t i = 0; i < this->m_mfccBuf.size(0); ++i) { - this->m_mfccBuf(i, mfccBufIdx) = mfcc[i]; - } - ++mfccBufIdx; - } - - /* Pad MFCC if needed by adding MFCC for zeros. */ - if (mfccBufIdx != this->m_numFeatureFrames) { - std::vector<int16_t> zerosWindow = std::vector<int16_t>(this->m_mfccWindowLen, 0); - std::vector<float> mfccZeros = this->m_mfcc.MfccCompute(zerosWindow); - - while (mfccBufIdx != this->m_numFeatureFrames) { - memcpy(&this->m_mfccBuf(0, mfccBufIdx), - mfccZeros.data(), sizeof(float) * m_numMfccFeats); - ++mfccBufIdx; - } - } - - /* Compute first and second order deltas from MFCCs. */ - AsrPreProcess::ComputeDeltas(this->m_mfccBuf, this->m_delta1Buf, this->m_delta2Buf); - - /* Standardize calculated features. */ - this->Standarize(); - - /* Quantise. */ - QuantParams quantParams = GetTensorQuantParams(this->m_inputTensor); - - if (0 == quantParams.scale) { - printf_err("Quantisation scale can't be 0\n"); - return false; - } - - switch(this->m_inputTensor->type) { - case kTfLiteUInt8: - return this->Quantise<uint8_t>( - tflite::GetTensorData<uint8_t>(this->m_inputTensor), this->m_inputTensor->bytes, - quantParams.scale, quantParams.offset); - case kTfLiteInt8: - return this->Quantise<int8_t>( - tflite::GetTensorData<int8_t>(this->m_inputTensor), this->m_inputTensor->bytes, - quantParams.scale, quantParams.offset); - default: - printf_err("Unsupported tensor type %s\n", - TfLiteTypeGetName(this->m_inputTensor->type)); - } - - return false; - } - - bool AsrPreProcess::ComputeDeltas(Array2d<float>& mfcc, - Array2d<float>& delta1, - Array2d<float>& delta2) - { - const std::vector <float> delta1Coeffs = - {6.66666667e-02, 5.00000000e-02, 3.33333333e-02, - 1.66666667e-02, -3.46944695e-18, -1.66666667e-02, - -3.33333333e-02, -5.00000000e-02, -6.66666667e-02}; - - const std::vector <float> delta2Coeffs = - {0.06060606, 0.01515152, -0.01731602, - -0.03679654, -0.04329004, -0.03679654, - -0.01731602, 0.01515152, 0.06060606}; - - if (delta1.size(0) == 0 || delta2.size(0) != delta1.size(0) || - mfcc.size(0) == 0 || mfcc.size(1) == 0) { - return false; - } - - /* Get the middle index; coeff vec len should always be odd. */ - const size_t coeffLen = delta1Coeffs.size(); - const size_t fMidIdx = (coeffLen - 1)/2; - const size_t numFeatures = mfcc.size(0); - const size_t numFeatVectors = mfcc.size(1); - - /* Iterate through features in MFCC vector. */ - for (size_t i = 0; i < numFeatures; ++i) { - /* For each feature, iterate through time (t) samples representing feature evolution and - * calculate d/dt and d^2/dt^2, using 1D convolution with differential kernels. - * Convolution padding = valid, result size is `time length - kernel length + 1`. - * The result is padded with 0 from both sides to match the size of initial time samples data. - * - * For the small filter, conv1D implementation as a simple loop is efficient enough. - * Filters of a greater size would need CMSIS-DSP functions to be used, like arm_fir_f32. - */ - - for (size_t j = fMidIdx; j < numFeatVectors - fMidIdx; ++j) { - float d1 = 0; - float d2 = 0; - const size_t mfccStIdx = j - fMidIdx; - - for (size_t k = 0, m = coeffLen - 1; k < coeffLen; ++k, --m) { - - d1 += mfcc(i,mfccStIdx + k) * delta1Coeffs[m]; - d2 += mfcc(i,mfccStIdx + k) * delta2Coeffs[m]; - } - - delta1(i,j) = d1; - delta2(i,j) = d2; - } - } - - return true; - } - - void AsrPreProcess::StandardizeVecF32(Array2d<float>& vec) - { - auto mean = math::MathUtils::MeanF32(vec.begin(), vec.totalSize()); - auto stddev = math::MathUtils::StdDevF32(vec.begin(), vec.totalSize(), mean); - - debug("Mean: %f, Stddev: %f\n", mean, stddev); - if (stddev == 0) { - std::fill(vec.begin(), vec.end(), 0); - } else { - const float stddevInv = 1.f/stddev; - const float normalisedMean = mean/stddev; - - auto NormalisingFunction = [=](float& value) { - value = value * stddevInv - normalisedMean; - }; - std::for_each(vec.begin(), vec.end(), NormalisingFunction); - } - } - - void AsrPreProcess::Standarize() - { - AsrPreProcess::StandardizeVecF32(this->m_mfccBuf); - AsrPreProcess::StandardizeVecF32(this->m_delta1Buf); - AsrPreProcess::StandardizeVecF32(this->m_delta2Buf); - } - - float AsrPreProcess::GetQuantElem( - const float elem, - const float quantScale, - const int quantOffset, - const float minVal, - const float maxVal) - { - float val = std::round((elem/quantScale) + quantOffset); - return std::min<float>(std::max<float>(val, minVal), maxVal); - } - -} /* namespace app */ -} /* namespace arm */
\ No newline at end of file diff --git a/source/use_case/kws_asr/usecase.cmake b/source/use_case/kws_asr/usecase.cmake index 40df4d7..59ef450 100644 --- a/source/use_case/kws_asr/usecase.cmake +++ b/source/use_case/kws_asr/usecase.cmake @@ -14,6 +14,8 @@ # See the License for the specific language governing permissions and # limitations under the License. #---------------------------------------------------------------------------- +# Append the APIs to use for this use case +list(APPEND ${use_case}_API_LIST "kws" "asr") USER_OPTION(${use_case}_FILE_PATH "Directory with WAV files, or path to a single WAV file, to use in the evaluation application." ${CMAKE_CURRENT_SOURCE_DIR}/resources/${use_case}/samples/ @@ -145,4 +147,4 @@ generate_audio_code(${${use_case}_FILE_PATH} ${SRC_GEN_DIR} ${INC_GEN_DIR} ${${use_case}_AUDIO_OFFSET} ${${use_case}_AUDIO_DURATION} ${${use_case}_AUDIO_RES_TYPE} - ${${use_case}_AUDIO_MIN_SAMPLES})
\ No newline at end of file + ${${use_case}_AUDIO_MIN_SAMPLES}) |