Opensource ML embedded evaluation kit
Change-Id: I12e807f19f5cacad7cef82572b6dd48252fd61fd
diff --git a/tests/use_case/asr/AsrClassifierTests.cc b/tests/use_case/asr/AsrClassifierTests.cc
new file mode 100644
index 0000000..7c71912
--- /dev/null
+++ b/tests/use_case/asr/AsrClassifierTests.cc
@@ -0,0 +1,98 @@
+/*
+ * 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 "Wav2LetterModel.hpp"
+
+#include <catch.hpp>
+
+TEST_CASE("Test invalid classifier")
+{
+ TfLiteTensor* outputTens = nullptr;
+ std::vector <arm::app::ClassificationResult> resultVec;
+ arm::app::AsrClassifier classifier;
+
+ REQUIRE(!classifier.GetClassificationResults(outputTens, resultVec, {}, 1));
+}
+
+
+TEST_CASE("Test valid classifier UINT8") {
+ const int dimArray[] = {4, 1, 1, 246, 29};
+ std::vector <std::string> labels(29);
+ std::vector <uint8_t> outputVec(7134);
+ TfLiteIntArray* dims= tflite::testing::IntArrayFromInts(dimArray);
+ TfLiteTensor tfTensor = tflite::testing::CreateQuantizedTensor(
+ outputVec.data(), dims, 1, 0, "test");
+ TfLiteTensor* outputTensor = &tfTensor;
+ std::vector <arm::app::ClassificationResult> resultVec;
+ arm::app::AsrClassifier classifier;
+
+ REQUIRE(classifier.GetClassificationResults(outputTensor, resultVec, labels, 1));
+ REQUIRE(246 == resultVec.size());
+}
+
+
+TEST_CASE("Get classification results") {
+ const int dimArray[] = {4, 1, 1, 10, 15};
+ std::vector <std::string> labels(15);
+ std::vector<uint8_t> outputVec(150, static_cast<uint8_t>(1));
+ TfLiteIntArray* dims= tflite::testing::IntArrayFromInts(dimArray);
+ TfLiteTensor tfTensor = tflite::testing::CreateQuantizedTensor(
+ outputVec.data(), dims, 1, 0, "test");
+ TfLiteTensor* outputTensor = &tfTensor;
+
+ std::vector <arm::app::ClassificationResult> resultVec(10);
+
+ /* set the top five results: */
+ std::vector<std::pair<uint32_t, std::pair<uint32_t, uint8_t>>> selectedResults {
+ {0, {3, 23}},
+ {0, {9, 15}},
+ {1, {5, 24}},
+ {1, {7, 4}},
+ {2, {9, 5}},
+ {3, {8, 6}},
+ {4, {13, 10}},
+ {4, {6, 18}},
+ {5, {3, 15}},
+ {5, {4, 115}},
+ {6, {6, 25}},
+ {7, {1, 7}},
+ {8, {11, 9}},
+ {9, {1, 10}}
+ };
+
+ const uint32_t nCols = outputTensor->dims->data[arm::app::Wav2LetterModel::ms_outputColsIdx];
+ for (size_t i = 0; i < selectedResults.size(); ++i) {
+ uint32_t rIndex = selectedResults[i].first;
+ uint32_t cIndex = selectedResults[i].second.first;
+ uint8_t value = selectedResults[i].second.second;
+ outputVec[rIndex * nCols + cIndex] = value;
+ }
+
+ arm::app::AsrClassifier classifier;
+
+ REQUIRE(classifier.GetClassificationResults(outputTensor, resultVec, labels, 1));
+ REQUIRE(resultVec[0].m_labelIdx == 3);
+ REQUIRE(resultVec[1].m_labelIdx == 5);
+ REQUIRE(resultVec[2].m_labelIdx == 9);
+ REQUIRE(resultVec[3].m_labelIdx == 8);
+ REQUIRE(resultVec[4].m_labelIdx == 6);
+ REQUIRE(resultVec[5].m_labelIdx == 4);
+ REQUIRE(resultVec[6].m_labelIdx == 6);
+ REQUIRE(resultVec[7].m_labelIdx == 1);
+ REQUIRE(resultVec[8].m_labelIdx == 11);
+ REQUIRE(resultVec[9].m_labelIdx == 1);
+}
diff --git a/tests/use_case/asr/AsrFeaturesTests.cc b/tests/use_case/asr/AsrFeaturesTests.cc
new file mode 100644
index 0000000..9401f40
--- /dev/null
+++ b/tests/use_case/asr/AsrFeaturesTests.cc
@@ -0,0 +1,188 @@
+/*
+ * 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 "DataStructures.hpp"
+#include "AsrGoldenFeatures.hpp"
+#include "hal.h"
+#include "TensorFlowLiteMicro.hpp"
+#include "Wav2LetterPreprocess.hpp"
+
+#include <catch.hpp>
+#include <random>
+
+class TestPreprocess : public arm::app::audio::asr::Preprocess {
+public:
+ TestPreprocess()
+ : arm::app::audio::asr::Preprocess(0,0,0,0)
+ {}
+
+ bool ComputeDeltas(arm::app::Array2d<float>& mfcc,
+ arm::app::Array2d<float>& delta1,
+ arm::app::Array2d<float>& delta2)
+ {
+ return this->_ComputeDeltas(mfcc, delta1, delta2);
+ }
+
+ float GetMean(arm::app::Array2d<float>& vec)
+ {
+ return this->_GetMean(vec);
+ }
+
+ float GetStdDev(arm::app::Array2d<float>& vec, const float mean)
+ {
+ return this->_GetStdDev(vec, mean);
+ }
+
+ void NormaliseVec(arm::app::Array2d<float>& vec)
+ {
+ return this->_NormaliseVec(vec);
+ }
+};
+
+template<class T>
+void CheckOutputs(const std::vector<T> goldenOutput, std::vector<T> output)
+{
+ const size_t goldenSize = goldenOutput.size();
+ const size_t realSize = output.size();
+
+ REQUIRE(realSize == goldenSize);
+ REQUIRE_THAT(output, Catch::Approx( goldenOutput ).margin(0.0001));
+}
+template void CheckOutputs<float>(const std::vector<float> goldenOutput, std::vector<float> output);
+
+void populateBuffer(const float* input, size_t size, size_t numMfccFeats, std::vector<std::vector<float>>& buf)
+{
+ size_t time = 0;
+ for (size_t i = 0; i < size; ++i) {
+ if (i > 0 && i % numMfccFeats == 0) {
+ ++time;
+ }
+ float featureValue = *(input + i);
+ buf[i % numMfccFeats][time] = featureValue;
+ }
+}
+
+void populateArray2dWithVectorOfVector(std::vector<std::vector<float>> vec, arm::app::Array2d<float>& buf)
+{
+ for (size_t i = 0; i < vec.size(); ++i) {
+ for (size_t j = 0; j < vec[i].size(); ++j) {
+ buf(i, j) = vec[i][j];
+ }
+ }
+}
+
+TEST_CASE("Floating point asr features calculation", "[ASR]")
+{
+ TestPreprocess tp;
+
+ SECTION("First and second diff")
+ {
+ constexpr uint32_t numMfccFeats = 13;
+ constexpr uint32_t numFeatVectors = 296;
+
+ arm::app::Array2d<float> mfccBuf(numMfccFeats, numFeatVectors);
+ arm::app::Array2d<float> delta1Buf(numMfccFeats, numFeatVectors);
+ arm::app::Array2d<float> delta2Buf(numMfccFeats, numFeatVectors);
+
+ std::vector<std::vector<float>> goldenMfccBuf(numMfccFeats, std::vector<float>(numFeatVectors));
+ std::vector<std::vector<float>> goldenDelta1Buf(numMfccFeats, std::vector<float>(numFeatVectors));
+ std::vector<std::vector<float>> goldenDelta2Buf(numMfccFeats, std::vector<float>(numFeatVectors));
+
+ populateBuffer(golden_asr_mfcc, golden_asr_mfcc_len, numMfccFeats, goldenMfccBuf);
+ populateBuffer(golden_diff1_features, golden_diff1_len, numMfccFeats, goldenDelta1Buf);
+ populateBuffer(golden_diff2_features, golden_diff2_len, numMfccFeats, goldenDelta2Buf);
+
+ populateArray2dWithVectorOfVector(goldenMfccBuf, mfccBuf);
+ std::fill(delta1Buf.begin(), delta1Buf.end(), 0.f);
+ std::fill(delta2Buf.begin(), delta2Buf.end(), 0.f);
+
+ tp.ComputeDeltas(mfccBuf, delta1Buf, delta2Buf);
+
+ /* First 4 and last 4 values are different because we pad AFTER diff calculated. */
+ for (size_t i = 0; i < numMfccFeats; ++i) {
+ const float* start_goldenDelta1Buf = goldenDelta1Buf[i].data() + 4;
+ const float* start_delta1 = delta1Buf.begin() + i * delta1Buf.size(1) + 4;
+ std::vector<float> goldenDataDelta1(start_goldenDelta1Buf, start_goldenDelta1Buf + numFeatVectors - 8);
+ std::vector<float> tensorDataDelta1(start_delta1, start_delta1 + numFeatVectors - 8);
+
+ CheckOutputs<float>(goldenDataDelta1,tensorDataDelta1);
+
+ const float* start_goldenDelta2Buf = goldenDelta2Buf[i].data() + 4;
+ const float* start_delta2 = delta2Buf.begin() + i * delta2Buf.size(1) + 4;
+ std::vector<float> goldenDataDelta2(start_goldenDelta2Buf, start_goldenDelta2Buf + numFeatVectors - 8);
+ std::vector<float> tensorDataDelta2(start_delta2, start_delta2 + numFeatVectors - 8);
+
+ CheckOutputs<float>(goldenDataDelta2,tensorDataDelta2);
+ }
+
+ }
+
+ SECTION("Mean")
+ {
+ std::vector<std::vector<float>> mean1vec{{1, 2},
+ {-1, -2}};
+ arm::app::Array2d<float> mean1(2,2); /* {{1, 2},{-1, -2}} */
+ populateArray2dWithVectorOfVector(mean1vec, mean1);
+ REQUIRE(0 == Approx(tp.GetMean(mean1)));
+
+ arm::app::Array2d<float> mean2(2, 2);
+ std::fill(mean2.begin(), mean2.end(), 0.f);
+ REQUIRE(0 == Approx(tp.GetMean(mean2)));
+
+ arm::app::Array2d<float> mean3(3,3);
+ std::fill(mean3.begin(), mean3.end(), 1.f);
+ REQUIRE(1 == Approx(tp.GetMean(mean3)));
+ }
+
+ SECTION("Std")
+ {
+ arm::app::Array2d<float> std1(2, 2);
+ std::fill(std1.begin(), std1.end(), 0.f); /* {{0, 0}, {0, 0}} */
+ REQUIRE(0 == Approx(tp.GetStdDev(std1, 0)));
+
+ std::vector<std::vector<float>> std2vec{{1, 2, 3, 4, 5}, {6, 7, 8, 9, 0}};
+ arm::app::Array2d<float> std2(2,5);
+ populateArray2dWithVectorOfVector(std2vec, std2);
+ const float mean = tp.GetMean(std2);
+ REQUIRE(2.872281323 == Approx(tp.GetStdDev(std2, mean)));
+
+ arm::app::Array2d<float> std3(2,2);
+ std::fill(std3.begin(), std3.end(), 1.f); /* std3{{1, 1}, {1, 1}}; */
+ REQUIRE(0 == Approx(tp.GetStdDev(std3, 1)));
+ }
+
+ SECTION("Norm") {
+ auto checker = [&](arm::app::Array2d<float>& d, std::vector<float>& g) {
+ tp.NormaliseVec(d);
+ std::vector<float> d_vec(d.begin(), d.end());
+ REQUIRE_THAT(g, Catch::Approx(d_vec));
+ };
+
+ std::vector<std::vector<float>> norm0vec{{1, 1}, {1, 1}};
+ std::vector<float> goldenNorm0 {0, 0, 0, 0};
+ arm::app::Array2d<float> norm0(2, 2);
+ populateArray2dWithVectorOfVector(norm0vec, norm0);
+ checker(norm0, goldenNorm0);
+
+ std::vector<std::vector<float>> norm1vec{{1, 2, 3, 4, 5}, {6, 7, 8, 9, 0}};
+ std::vector<float> goldenNorm1 {
+ -1.218543592, -0.87038828, -0.522232968, -0.174077656, 0.174077656,
+ 0.522232968, 0.87038828, 1.218543592, 1.566698904, -1.566698904};
+ arm::app::Array2d<float> norm1(2, 5);
+ populateArray2dWithVectorOfVector(norm1vec, norm1);
+ checker(norm1, goldenNorm1);
+ }
+}
diff --git a/tests/use_case/asr/AsrTests.cc b/tests/use_case/asr/AsrTests.cc
new file mode 100644
index 0000000..09f82da
--- /dev/null
+++ b/tests/use_case/asr/AsrTests.cc
@@ -0,0 +1,18 @@
+/*
+ * 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.
+ */
+#define CATCH_CONFIG_MAIN
+#include <catch.hpp>
diff --git a/tests/use_case/asr/InferenceTestWav2Letter.cc b/tests/use_case/asr/InferenceTestWav2Letter.cc
new file mode 100644
index 0000000..1fa4092
--- /dev/null
+++ b/tests/use_case/asr/InferenceTestWav2Letter.cc
@@ -0,0 +1,105 @@
+/*
+ * 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 "hal.h"
+#include "TensorFlowLiteMicro.hpp"
+#include "Wav2LetterModel.hpp"
+#include "TestData_asr.hpp"
+
+#include <catch.hpp>
+#include <random>
+
+bool RunInference(arm::app::Model& model, const int8_t vec[], const size_t copySz)
+{
+ TfLiteTensor* inputTensor = model.GetInputTensor(0);
+ REQUIRE(inputTensor);
+
+ memcpy(inputTensor->data.data, vec, copySz);
+
+ return model.RunInference();
+}
+
+bool RunInferenceRandom(arm::app::Model& model)
+{
+ TfLiteTensor* inputTensor = model.GetInputTensor(0);
+ REQUIRE(inputTensor);
+
+ std::random_device rndDevice;
+ std::mt19937 mersenneGen{rndDevice()};
+ std::uniform_int_distribution<short> dist {-128, 127};
+
+ auto gen = [&dist, &mersenneGen](){
+ return dist(mersenneGen);
+ };
+
+ std::vector<int8_t> randomAudio(inputTensor->bytes);
+ std::generate(std::begin(randomAudio), std::end(randomAudio), gen);
+
+ REQUIRE(RunInference(model, randomAudio.data(), inputTensor->bytes));
+ return true;
+}
+
+/* Skip this test, Wav2LetterModel if not Vela optimized but only from ML-zoo will fail. */
+TEST_CASE("Running random inference with TensorFlow Lite Micro and Wav2LetterModel Int8", "[Wav2Letter][.]")
+{
+ arm::app::Wav2LetterModel model{};
+
+ REQUIRE_FALSE(model.IsInited());
+ REQUIRE(model.Init());
+ REQUIRE(model.IsInited());
+
+ REQUIRE(RunInferenceRandom(model));
+}
+
+template<typename T>
+void TestInference(const T* input_goldenFV, const T* output_goldenFV, arm::app::Model& model)
+{
+ TfLiteTensor* inputTensor = model.GetInputTensor(0);
+ REQUIRE(inputTensor);
+
+ REQUIRE(RunInference(model, input_goldenFV, inputTensor->bytes));
+
+ TfLiteTensor* outputTensor = model.GetOutputTensor(0);
+
+ REQUIRE(outputTensor);
+ REQUIRE(outputTensor->bytes == OFM_DATA_SIZE);
+ auto tensorData = tflite::GetTensorData<T>(outputTensor);
+ REQUIRE(tensorData);
+
+ for (size_t i = 0; i < outputTensor->bytes; i++) {
+ REQUIRE((int)tensorData[i] == (int)((T)output_goldenFV[i]));
+ }
+}
+
+TEST_CASE("Running inference with Tflu and Wav2LetterModel Int8", "[Wav2Letter][.]")
+{
+ for (uint32_t i = 0 ; i < NUMBER_OF_FM_FILES; ++i) {
+ auto input_goldenFV = get_ifm_data_array(i);;
+ auto output_goldenFV = get_ofm_data_array(i);
+
+ DYNAMIC_SECTION("Executing inference with re-init")
+ {
+ arm::app::Wav2LetterModel model{};
+
+ REQUIRE_FALSE(model.IsInited());
+ REQUIRE(model.Init());
+ REQUIRE(model.IsInited());
+
+ TestInference<int8_t>(input_goldenFV, output_goldenFV, model);
+
+ }
+ }
+}
\ No newline at end of file
diff --git a/tests/use_case/asr/MfccTests.cc b/tests/use_case/asr/MfccTests.cc
new file mode 100644
index 0000000..c70e53e
--- /dev/null
+++ b/tests/use_case/asr/MfccTests.cc
@@ -0,0 +1,170 @@
+/*
+ * 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 <algorithm>
+#include <catch.hpp>
+#include <limits>
+
+/* First 512 samples from itellyou.wav. */
+const std::vector<int16_t> testWav1 = std::vector<int16_t> {
+ -3,0,1,-1,2,3,-2,2,
+ 1,-2,0,3,-1,8,3,2,
+ -1,-1,2,7,3,5,6,6,
+ 6,12,5,6,3,3,5,4,
+ 4,6,7,7,7,3,7,2,
+ 8,4,4,2,-4,-1,-1,-4,
+ 2,1,-1,-4,0,-7,-6,-2,
+ -5,1,-5,-1,-7,-3,-3,-7,
+ 0,-3,3,-5,0,1,-2,-2,
+ -3,-3,-7,-3,-2,-6,-5,-8,
+ -2,-8,4,-9,-4,-9,-5,-5,
+ -3,-9,-3,-9,-1,-7,-4,1,
+ -3,2,-8,-4,-4,-5,1,-3,
+ -1,0,-1,-2,-3,-2,-4,-1,
+ 1,-1,3,0,3,2,0,0,
+ 0,-3,1,1,0,8,3,4,
+ 1,5,6,4,7,3,3,0,
+ 3,6,7,6,4,5,9,9,
+ 5,5,8,1,6,9,6,6,
+ 7,1,8,1,5,0,5,5,
+ 0,3,2,7,2,-3,3,0,
+ 3,0,0,0,2,0,-1,-1,
+ -2,-3,-8,0,1,0,-3,-3,
+ -3,-2,-3,-3,-4,-6,-2,-8,
+ -9,-4,-1,-5,-3,-3,-4,-3,
+ -6,3,0,-1,-2,-9,-4,-2,
+ 2,-1,3,-5,-5,-2,0,-2,
+ 0,-1,-3,1,-2,9,4,5,
+ 2,2,1,0,-6,-2,0,0,
+ 0,-1,4,-4,3,-7,-1,5,
+ -6,-1,-5,4,3,9,-2,1,
+ 3,0,0,-2,1,2,1,1,
+ 0,3,2,-1,3,-3,7,0,
+ 0,3,2,2,-2,3,-2,2,
+ -3,4,-1,-1,-5,-1,-3,-2,
+ 1,-1,3,2,4,1,2,-2,
+ 0,2,7,0,8,-3,6,-3,
+ 6,1,2,-3,-1,-1,-1,1,
+ -2,2,1,2,0,-2,3,-2,
+ 3,-2,1,0,-3,-1,-2,-4,
+ -6,-5,-8,-1,-4,0,-3,-1,
+ -1,-1,0,-2,-3,-7,-1,0,
+ 1,5,0,5,1,1,-3,0,
+ -6,3,-8,4,-8,6,-6,1,
+ -6,-2,-5,-6,0,-5,4,-1,
+ 4,-2,1,2,1,0,-2,0,
+ 0,2,-2,2,-5,2,0,-2,
+ 1,-2,0,5,1,0,1,5,
+ 0,8,3,2,2,0,5,-2,
+ 3,1,0,1,0,-2,-1,-3,
+ 1,-1,3,0,3,0,-2,-1,
+ -4,-4,-4,-1,-4,-4,-3,-6,
+ -3,-7,-3,-1,-2,0,-5,-4,
+ -7,-3,-2,-2,1,2,2,8,
+ 5,4,2,4,3,5,0,3,
+ 3,6,4,2,2,-2,4,-2,
+ 3,3,2,1,1,4,-5,2,
+ -3,0,-1,1,-2,2,5,1,
+ 4,2,3,1,-1,1,0,6,
+ 0,-2,-1,1,-1,2,-5,-1,
+ -5,-1,-6,-3,-3,2,4,0,
+ -1,-5,3,-4,-1,-3,-4,1,
+ -4,1,-1,-1,0,-5,-4,-2,
+ -1,-1,-3,-7,-3,-3,4,4,
+};
+
+const std::vector<int16_t> testWav2 = std::vector<int16_t> (512, 0);
+
+/* Golden mfcc output for testwav1. */
+const std::vector<float> golden_mfcc_output_testWav1 {
+ -835.24603, 21.010452, 18.699404, 7.4338417, 19.028961, -5.401735, 6.4761047, -11.400679,
+ 8.392709, 12.202361, 8.403276, -13.508412, -18.307348
+};
+
+/* Golden mfcc output for the all zero wav. */
+const std::vector<float> golden_mfcc_output_testWav2 {
+ -1131.37085, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
+};
+
+
+arm::app::audio::Wav2LetterMFCC GetMFCCInstance()
+{
+ const auto sampFreq = arm::app::audio::Wav2LetterMFCC::ms_defaultSamplingFreq;
+ const auto frameLenMs = 32;
+ const auto numMfccFeats = 13;
+ const auto frameLenSamples = sampFreq * frameLenMs * 0.001;
+ return arm::app::audio::Wav2LetterMFCC(numMfccFeats, frameLenSamples);
+}
+
+template <class T>
+void TestQuantisedMFCC()
+{
+ const auto quantScale = 0.1410219967365265;
+ const auto quantOffset = 11;
+ std::vector<T> mfccOutput = GetMFCCInstance().MfccComputeQuant<T>(testWav1, quantScale, quantOffset);
+
+ long min_val = std::numeric_limits<T>::min();
+ long max_val = std::numeric_limits<T>::max();
+
+ for (size_t i = 0; i < golden_mfcc_output_testWav1.size(); i++){
+ long TestWavMfcc = (std::lround((golden_mfcc_output_testWav1[i] / quantScale) + quantOffset));
+ T quantizedTestWavMfcc = static_cast<T>(std::max(min_val, std::min(TestWavMfcc, max_val)));
+
+ REQUIRE(quantizedTestWavMfcc == Approx(mfccOutput[i]).margin(2));
+ }
+}
+
+template void TestQuantisedMFCC<int8_t>();
+template void TestQuantisedMFCC<uint8_t>();
+template void TestQuantisedMFCC<int16_t>();
+
+TEST_CASE("MFCC calculation")
+{
+ hal_platform platform;
+ data_acq_module dataAcq;
+ data_psn_module dataPsn;
+ platform_timer timer;
+
+ /* Initialise the HAL and platform. */
+ hal_init(&platform, &dataAcq, &dataPsn, &timer);
+ hal_platform_init(&platform);
+
+ SECTION("FP32")
+ {
+ auto mfccOutput = GetMFCCInstance().MfccCompute(testWav1);
+ REQUIRE_THAT( mfccOutput, Catch::Approx( golden_mfcc_output_testWav1 ).margin(0.3) );
+
+ auto mfccOutput2 = GetMFCCInstance().MfccCompute(testWav2);
+ REQUIRE_THAT( mfccOutput2, Catch::Approx( golden_mfcc_output_testWav2 ).margin(0.001) );
+ }
+
+ SECTION("int8_t")
+ {
+ TestQuantisedMFCC<int8_t>();
+ }
+
+ SECTION("uint8_t")
+ {
+ TestQuantisedMFCC<uint8_t>();
+ }
+
+ SECTION("int16_t")
+ {
+ TestQuantisedMFCC<int16_t>();
+ }
+}
diff --git a/tests/use_case/asr/OutputDecodeTests.cc b/tests/use_case/asr/OutputDecodeTests.cc
new file mode 100644
index 0000000..22153f3
--- /dev/null
+++ b/tests/use_case/asr/OutputDecodeTests.cc
@@ -0,0 +1,67 @@
+/*
+ * 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"
+
+#include "catch.hpp"
+
+TEST_CASE("Running output decode on test vector") {
+
+ std::vector<arm::app::ClassificationResult> vecResult(20);
+ /* Number of test inputs. */
+ const size_t numStrings = 8;
+
+ /* The test inputs. */
+ std::string testText[numStrings][20]
+ {
+ {"a", "b", "c", "d", "e", "f", "g", "g", "g", " ", "h", "h", "i", " ", " ", "j", "k", "\'", "\'", "l"}, /* initial */
+ {" ", "b", "c", "d", "e", "f", "g", "g", "g", " ", "h", "h", "i", " ", " ", "j", "k", "\'", "\'", " "}, /* space start and end */
+ {"\'", "b", "c", "d", "e", "f", "g", "g", "g", " ", "h", "h", "i", " ", " ", "j", "k", "\'", "l", "\'"}, /* apostrophe start and end */
+ {"a", "a", "c", "d", "e", "f", "g", "g", "g", " ", "h", "h", "i", " ", " ", "j", "k", "\'", "l", "l"}, /* Double start and end */
+ {"a", "b", "c", "d", "e", "f", "g", "g", "o", "$", "o", "h", "i", " ", " ", "j", "k", "\'", "\'", "l"}, /* Legit double character */
+ {"a", "$", "a", "d", "e", "f", "g", "g", "o", "$", "o", "h", "i", " ", " ", "j", "k", "l", "$", "l"}, /* Legit double character start and end */
+ {"$", "a", "b", "d", "e", "f", "g", "g", "o", "$", "o", "h", "i", " ", " ", "j", "k", "l", "$", "$"}, /* $$ */
+ {"$", "a", "b", "d", "e", "f", "g", "g", "o", "$", "o", "h", "i", " ", " ", "j", "k", "l", "l", "l"}
+ };
+
+ /* The golden outputs for the above test inputs. */
+ std::string expectedOutput[numStrings] =
+ {
+ {"abcdefg hi jk\'l"},
+ {" bcdefg hi jk\' "},
+ {"\'bcdefg hi jk\'l\'"},
+ {"acdefg hi jk\'l"},
+ {"abcdefgoohi jk\'l"},
+ {"aadefgoohi jkll"},
+ {"abdefgoohi jkl"},
+ {"abdefgoohi jkl"}
+ };
+
+ /*For each test input. */
+ for (size_t h = 0; h < numStrings; ++h)
+ {
+ /* Generate fake vecResults.m_label to mimic AsrClassifier output containing the testText. */
+ for (size_t i = 0; i < 20; i++)
+ {
+ vecResult[i].m_label = testText[h][i];
+ }
+ /* Call function with fake vecResults and save returned string into 'buff'. */
+ std::string buff = arm::app::audio::asr::DecodeOutput(vecResult);
+
+ /* Check that the string returned from the function matches the expected output given above. */
+ REQUIRE(buff.compare(expectedOutput[h]) == 0);
+ }
+}
\ No newline at end of file
diff --git a/tests/use_case/asr/Wav2LetterPostprocessingTest.cc b/tests/use_case/asr/Wav2LetterPostprocessingTest.cc
new file mode 100644
index 0000000..9ed2e1b
--- /dev/null
+++ b/tests/use_case/asr/Wav2LetterPostprocessingTest.cc
@@ -0,0 +1,199 @@
+/*
+ * 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 "Wav2LetterPostprocess.hpp"
+#include "Wav2LetterModel.hpp"
+
+#include <algorithm>
+#include <catch.hpp>
+#include <limits>
+
+template <typename T>
+static TfLiteTensor GetTestTensor(
+ std::vector <int>& shape,
+ T initVal,
+ std::vector<T>& vectorBuf)
+{
+ REQUIRE(0 != shape.size());
+
+ shape.insert(shape.begin(), shape.size());
+ uint32_t sizeInBytes = sizeof(T);
+ for (size_t i = 1; i < shape.size(); ++i) {
+ sizeInBytes *= shape[i];
+ }
+
+ /* Allocate mem. */
+ vectorBuf = std::vector<T>(sizeInBytes, initVal);
+ TfLiteIntArray* dims = tflite::testing::IntArrayFromInts(shape.data());
+ return tflite::testing::CreateQuantizedTensor(
+ vectorBuf.data(), dims,
+ 1, 0, "test-tensor");
+}
+
+TEST_CASE("Checking return value")
+{
+ SECTION("Mismatched post processing parameters and tensor size")
+ {
+ const uint32_t ctxLen = 5;
+ const uint32_t innerLen = 3;
+ arm::app::audio::asr::Postprocess post{ctxLen, innerLen, 0};
+
+ std::vector <int> tensorShape = {1, 1, 1, 13};
+ std::vector <int8_t> tensorVec;
+ TfLiteTensor tensor = GetTestTensor<int8_t>(
+ tensorShape, 100, tensorVec);
+ REQUIRE(false == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, false));
+ }
+
+ SECTION("Post processing succeeds")
+ {
+ const uint32_t ctxLen = 5;
+ const uint32_t innerLen = 3;
+ arm::app::audio::asr::Postprocess post{ctxLen, innerLen, 0};
+
+ std::vector <int> tensorShape = {1, 1, 13, 1};
+ std::vector <int8_t> tensorVec;
+ TfLiteTensor tensor = GetTestTensor<int8_t>(
+ tensorShape, 100, tensorVec);
+
+ /* Copy elements to compare later. */
+ std::vector <int8_t> originalVec = tensorVec;
+
+ /* This step should not erase anything. */
+ REQUIRE(true == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, false));
+ }
+}
+
+
+TEST_CASE("Postprocessing - erasing required elements")
+{
+ constexpr uint32_t ctxLen = 5;
+ constexpr uint32_t innerLen = 3;
+ constexpr uint32_t nRows = 2*ctxLen + innerLen;
+ constexpr uint32_t nCols = 10;
+ constexpr uint32_t blankTokenIdx = nCols - 1;
+ std::vector <int> tensorShape = {1, 1, nRows, nCols};
+
+ SECTION("First and last iteration")
+ {
+ arm::app::audio::asr::Postprocess post{ctxLen, innerLen, blankTokenIdx};
+
+ std::vector <int8_t> tensorVec;
+ TfLiteTensor tensor = GetTestTensor<int8_t>(
+ tensorShape, 100, tensorVec);
+
+ /* Copy elements to compare later. */
+ std::vector <int8_t> originalVec = tensorVec;
+
+ /* This step should not erase anything. */
+ REQUIRE(true == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, true));
+ REQUIRE(originalVec == tensorVec);
+ }
+
+ SECTION("Right context erase")
+ {
+ arm::app::audio::asr::Postprocess post{ctxLen, innerLen, blankTokenIdx};
+
+ std::vector <int8_t> tensorVec;
+ TfLiteTensor tensor = GetTestTensor<int8_t>(
+ tensorShape, 100, tensorVec);
+
+ /* Copy elements to compare later. */
+ std::vector <int8_t> originalVec = tensorVec;
+
+ /* This step should erase the right context only. */
+ REQUIRE(true == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, false));
+ REQUIRE(originalVec != tensorVec);
+
+ /* The last ctxLen * 10 elements should be gone. */
+ for (size_t i = 0; i < ctxLen; ++i) {
+ for (size_t j = 0; j < nCols; ++j) {
+ /* Check right context elements are zeroed. */
+ if (j == blankTokenIdx) {
+ CHECK(tensorVec[(ctxLen + innerLen) * nCols + i*nCols + j] == 1);
+ } else {
+ CHECK(tensorVec[(ctxLen + innerLen) * nCols + i*nCols + j] == 0);
+ }
+
+ /* Check left context is preserved. */
+ CHECK(tensorVec[i*nCols + j] == originalVec[i*nCols + j]);
+ }
+ }
+
+ /* Check inner elements are preserved. */
+ for (size_t i = ctxLen * nCols; i < (ctxLen + innerLen) * nCols; ++i) {
+ CHECK(tensorVec[i] == originalVec[i]);
+ }
+ }
+
+ SECTION("Left and right context erase")
+ {
+ arm::app::audio::asr::Postprocess post{ctxLen, innerLen, blankTokenIdx};
+
+ std::vector <int8_t> tensorVec;
+ TfLiteTensor tensor = GetTestTensor<int8_t>(
+ tensorShape, 100, tensorVec);
+
+ /* Copy elements to compare later. */
+ std::vector <int8_t> originalVec = tensorVec;
+
+ /* This step should erase right context. */
+ REQUIRE(true == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, false));
+
+ /* Calling it the second time should erase the left context. */
+ REQUIRE(true == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, false));
+
+ REQUIRE(originalVec != tensorVec);
+
+ /* The first and last ctxLen * 10 elements should be gone. */
+ for (size_t i = 0; i < ctxLen; ++i) {
+ for (size_t j = 0; j < nCols; ++j) {
+ /* Check left and right context elements are zeroed. */
+ if (j == blankTokenIdx) {
+ CHECK(tensorVec[(ctxLen + innerLen) * nCols + i*nCols + j] == 1);
+ CHECK(tensorVec[i*nCols + j] == 1);
+ } else {
+ CHECK(tensorVec[(ctxLen + innerLen) * nCols + i*nCols + j] == 0);
+ CHECK(tensorVec[i*nCols + j] == 0);
+ }
+ }
+ }
+
+ /* Check inner elements are preserved. */
+ for (size_t i = ctxLen * nCols; i < (ctxLen + innerLen) * nCols; ++i) {
+ /* Check left context is preserved. */
+ CHECK(tensorVec[i] == originalVec[i]);
+ }
+ }
+
+ SECTION("Try left context erase")
+ {
+ /* Should not be able to erase the left context if it is the first iteration. */
+ arm::app::audio::asr::Postprocess post{ctxLen, innerLen, blankTokenIdx};
+
+ std::vector <int8_t> tensorVec;
+ TfLiteTensor tensor = GetTestTensor<int8_t>(
+ tensorShape, 100, tensorVec);
+
+ /* Copy elements to compare later. */
+ std::vector <int8_t> originalVec = tensorVec;
+
+ /* Calling it the second time should erase the left context. */
+ REQUIRE(true == post.Invoke(&tensor, arm::app::Wav2LetterModel::ms_outputRowsIdx, true));
+
+ REQUIRE(originalVec == tensorVec);
+ }
+}
diff --git a/tests/use_case/asr/Wav2LetterPreprocessingTest.cc b/tests/use_case/asr/Wav2LetterPreprocessingTest.cc
new file mode 100644
index 0000000..1391011
--- /dev/null
+++ b/tests/use_case/asr/Wav2LetterPreprocessingTest.cc
@@ -0,0 +1,152 @@
+/*
+ * 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 "Wav2LetterPreprocess.hpp"
+
+#include <limits>
+#include <algorithm>
+#include <catch.hpp>
+
+constexpr uint32_t numMfccFeatures = 13;
+constexpr uint32_t numMfccVectors = 10;
+
+/* Test vector output: generated using test-asr-preprocessing.py. */
+int8_t expectedResult[numMfccVectors][numMfccFeatures * 3] = {
+ /* Feature vec 0. */
+ -32, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11, /* MFCCs. */
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, /* Delta 1. */
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, /* Delta 2. */
+
+ /* Feature vec 1. */
+ -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
+
+ /* Feature vec 2. */
+ -31, 4, -9, -9, -10, -10, -11, -11, -11, -11, -12, -12, -12,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
+
+ /* Feature vec 3. */
+ -31, 4, -9, -9, -10, -10, -11, -11, -11, -11, -11, -12, -12,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
+
+ /* Feature vec 4 : this should have valid delta 1 and delta 2. */
+ -31, 4, -9, -9, -10, -10, -11, -11, -11, -11, -11, -12, -12,
+ -38, -29, -9, 1, -2, -7, -8, -8, -12, -16, -14, -5, 5,
+ -68, -50, -13, 5, 0, -9, -9, -8, -13, -20, -19, -3, 15,
+
+ /* Feature vec 5 : this should have valid delta 1 and delta 2. */
+ -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -11, -12, -12,
+ -62, -45, -11, 5, 0, -8, -9, -8, -12, -19, -17, -3, 13,
+ -27, -22, -13, -9, -11, -12, -12, -11, -11, -13, -13, -10, -6,
+
+ /* Feature vec 6. */
+ -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
+
+ /* Feature vec 7. */
+ -32, 4, -9, -8, -10, -10, -11, -11, -11, -12, -12, -11, -11,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
+
+ /* Feature vec 8. */
+ -32, 4, -9, -8, -10, -10, -11, -11, -11, -12, -12, -11, -11,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10,
+
+ /* Feature vec 9. */
+ -31, 4, -9, -8, -10, -10, -11, -11, -11, -11, -12, -11, -11,
+ -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11, -11,
+ -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10, -10
+};
+
+void PopulateTestWavVector(std::vector<int16_t>& vec)
+{
+ constexpr int int16max = std::numeric_limits<int16_t>::max();
+ int val = 0;
+ for (size_t i = 0; i < vec.size(); ++i, ++val) {
+
+ /* We want a differential filter response from both - order 1
+ * and 2 => Don't have a linear signal here - we use a signal
+ * using squares for example. Alternate sign flips might work
+ * just as well and will be computationally less work! */
+ int valsq = val * val;
+ if (valsq > int16max) {
+ val = 0;
+ valsq = 0;
+ }
+ vec[i] = valsq;
+ }
+}
+
+TEST_CASE("Preprocessing calculation INT8")
+{
+ /* Initialise the HAL and platform. */
+ hal_platform platform;
+ data_acq_module data_acq;
+ data_psn_module data_psn;
+ platform_timer timer;
+ hal_init(&platform, &data_acq, &data_psn, &timer);
+ hal_platform_init(&platform);
+
+ /* Constants. */
+ const uint32_t windowLen = 512;
+ const uint32_t windowStride = 160;
+ const int dimArray[] = {3, 1, numMfccFeatures * 3, numMfccVectors};
+ const float quantScale = 0.1410219967365265;
+ const int quantOffset = -11;
+
+ /* Test wav memory. */
+ std::vector <int16_t> testWav((windowStride * numMfccVectors) +
+ (windowLen - windowStride));
+
+ /* Populate with dummy input. */
+ PopulateTestWavVector(testWav);
+
+ /* Allocate mem for tensor. */
+ std::vector<int8_t> tensorVec(dimArray[1]*dimArray[2]*dimArray[3]);
+
+ /* Initialise dimensions and the test tensor. */
+ TfLiteIntArray* dims= tflite::testing::IntArrayFromInts(dimArray);
+ TfLiteTensor tensor = tflite::testing::CreateQuantizedTensor(
+ tensorVec.data(), dims, quantScale, quantOffset, "preprocessedInput");
+
+ /* Initialise pre-processing module. */
+ arm::app::audio::asr::Preprocess prep{
+ numMfccFeatures, windowLen, windowStride, numMfccVectors};
+
+ /* Invoke pre-processing. */
+ REQUIRE(prep.Invoke(testWav.data(), testWav.size(), &tensor));
+
+ /* Wrap the tensor with a std::vector for ease. */
+ int8_t * tensorData = tflite::GetTensorData<int8_t>(&tensor);
+ std::vector <int8_t> vecResults =
+ std::vector<int8_t>(tensorData, tensorData + tensor.bytes);
+
+ /* Check sizes. */
+ REQUIRE(vecResults.size() == sizeof(expectedResult));
+
+ /* Check that the elements have been calculated correctly. */
+ for (uint32_t j = 0; j < numMfccVectors; ++j) {
+ for (uint32_t i = 0; i < numMfccFeatures * 3; ++i) {
+ size_t tensorIdx = (j * numMfccFeatures * 3) + i;
+ CHECK(vecResults[tensorIdx] == expectedResult[j][i]);
+ }
+ }
+}