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 "AsrClassifier.hpp" |
| 18 | |
alexander | 31ae9f0 | 2022-02-10 16:15:54 +0000 | [diff] [blame] | 19 | #include "log_macros.h" |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 20 | #include "TensorFlowLiteMicro.hpp" |
| 21 | #include "Wav2LetterModel.hpp" |
| 22 | |
Richard Burton | b40ecf8 | 2022-04-22 16:14:57 +0100 | [diff] [blame] | 23 | namespace arm { |
| 24 | namespace app { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 25 | |
Richard Burton | b40ecf8 | 2022-04-22 16:14:57 +0100 | [diff] [blame] | 26 | template<typename T> |
| 27 | bool AsrClassifier::GetTopResults(TfLiteTensor* tensor, |
| 28 | std::vector<ClassificationResult>& vecResults, |
| 29 | const std::vector <std::string>& labels, double scale, double zeroPoint) |
| 30 | { |
| 31 | const uint32_t nElems = tensor->dims->data[Wav2LetterModel::ms_outputRowsIdx]; |
| 32 | const uint32_t nLetters = tensor->dims->data[Wav2LetterModel::ms_outputColsIdx]; |
alexander | c350cdc | 2021-04-29 20:36:09 +0100 | [diff] [blame] | 33 | |
Richard Burton | b40ecf8 | 2022-04-22 16:14:57 +0100 | [diff] [blame] | 34 | if (nLetters != labels.size()) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 35 | printf("Output size doesn't match the labels' size\n"); |
| 36 | return false; |
| 37 | } |
| 38 | |
Richard Burton | b40ecf8 | 2022-04-22 16:14:57 +0100 | [diff] [blame] | 39 | /* NOTE: tensor's size verification against labels should be |
| 40 | * checked by the calling/public function. */ |
| 41 | if (nLetters < 1) { |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 42 | return false; |
| 43 | } |
| 44 | |
Richard Burton | b40ecf8 | 2022-04-22 16:14:57 +0100 | [diff] [blame] | 45 | /* Final results' container. */ |
| 46 | vecResults = std::vector<ClassificationResult>(nElems); |
| 47 | |
| 48 | T* tensorData = tflite::GetTensorData<T>(tensor); |
| 49 | |
| 50 | /* Get the top 1 results. */ |
| 51 | for (uint32_t i = 0, row = 0; i < nElems; ++i, row+=nLetters) { |
| 52 | std::pair<T, uint32_t> top_1 = std::make_pair(tensorData[row + 0], 0); |
| 53 | |
| 54 | for (uint32_t j = 1; j < nLetters; ++j) { |
| 55 | if (top_1.first < tensorData[row + j]) { |
| 56 | top_1.first = tensorData[row + j]; |
| 57 | top_1.second = j; |
| 58 | } |
| 59 | } |
| 60 | |
| 61 | double score = static_cast<int> (top_1.first); |
| 62 | vecResults[i].m_normalisedVal = scale * (score - zeroPoint); |
| 63 | vecResults[i].m_label = labels[top_1.second]; |
| 64 | vecResults[i].m_labelIdx = top_1.second; |
| 65 | } |
| 66 | |
alexander | 3c79893 | 2021-03-26 21:42:19 +0000 | [diff] [blame] | 67 | return true; |
Richard Burton | b40ecf8 | 2022-04-22 16:14:57 +0100 | [diff] [blame] | 68 | } |
| 69 | template bool AsrClassifier::GetTopResults<uint8_t>(TfLiteTensor* tensor, |
| 70 | std::vector<ClassificationResult>& vecResults, |
| 71 | const std::vector <std::string>& labels, |
| 72 | double scale, double zeroPoint); |
| 73 | template bool AsrClassifier::GetTopResults<int8_t>(TfLiteTensor* tensor, |
| 74 | std::vector<ClassificationResult>& vecResults, |
| 75 | const std::vector <std::string>& labels, |
| 76 | double scale, double zeroPoint); |
| 77 | |
| 78 | bool AsrClassifier::GetClassificationResults( |
| 79 | TfLiteTensor* outputTensor, |
| 80 | std::vector<ClassificationResult>& vecResults, |
| 81 | const std::vector <std::string>& labels, uint32_t topNCount, bool use_softmax) |
| 82 | { |
| 83 | UNUSED(use_softmax); |
| 84 | vecResults.clear(); |
| 85 | |
| 86 | constexpr int minTensorDims = static_cast<int>( |
| 87 | (Wav2LetterModel::ms_outputRowsIdx > Wav2LetterModel::ms_outputColsIdx)? |
| 88 | Wav2LetterModel::ms_outputRowsIdx : Wav2LetterModel::ms_outputColsIdx); |
| 89 | |
| 90 | constexpr uint32_t outColsIdx = Wav2LetterModel::ms_outputColsIdx; |
| 91 | |
| 92 | /* Sanity checks. */ |
| 93 | if (outputTensor == nullptr) { |
| 94 | printf_err("Output vector is null pointer.\n"); |
| 95 | return false; |
| 96 | } else if (outputTensor->dims->size < minTensorDims) { |
| 97 | printf_err("Output tensor expected to be %dD\n", minTensorDims); |
| 98 | return false; |
| 99 | } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) < topNCount) { |
| 100 | printf_err("Output vectors are smaller than %" PRIu32 "\n", topNCount); |
| 101 | return false; |
| 102 | } else if (static_cast<uint32_t>(outputTensor->dims->data[outColsIdx]) != labels.size()) { |
| 103 | printf("Output size doesn't match the labels' size\n"); |
| 104 | return false; |
| 105 | } |
| 106 | |
| 107 | if (topNCount != 1) { |
| 108 | warn("TopNCount value ignored in this implementation\n"); |
| 109 | } |
| 110 | |
| 111 | /* To return the floating point values, we need quantization parameters. */ |
| 112 | QuantParams quantParams = GetTensorQuantParams(outputTensor); |
| 113 | |
| 114 | bool resultState; |
| 115 | |
| 116 | switch (outputTensor->type) { |
| 117 | case kTfLiteUInt8: |
| 118 | resultState = this->GetTopResults<uint8_t>( |
| 119 | outputTensor, vecResults, |
| 120 | labels, quantParams.scale, |
| 121 | quantParams.offset); |
| 122 | break; |
| 123 | case kTfLiteInt8: |
| 124 | resultState = this->GetTopResults<int8_t>( |
| 125 | outputTensor, vecResults, |
| 126 | labels, quantParams.scale, |
| 127 | quantParams.offset); |
| 128 | break; |
| 129 | default: |
| 130 | printf_err("Tensor type %s not supported by classifier\n", |
| 131 | TfLiteTypeGetName(outputTensor->type)); |
| 132 | return false; |
| 133 | } |
| 134 | |
| 135 | if (!resultState) { |
| 136 | printf_err("Failed to get sorted set\n"); |
| 137 | return false; |
| 138 | } |
| 139 | |
| 140 | return true; |
| 141 | } |
| 142 | |
| 143 | } /* namespace app */ |
| 144 | } /* namespace arm */ |