MLECO-3164: Additional refactoring of KWS API
Part 1
* Add KwsClassifier
* KwsPostProcess can now be told to average results
* Averaging is handlded by KwsClassifier
* Current sliding window index is now an argument of DoPreProcess
Change-Id: I07626da595ad1cbd982e8366f0d1bb56d1040459
diff --git a/source/application/api/use_case/kws/include/KwsClassifier.hpp b/source/application/api/use_case/kws/include/KwsClassifier.hpp
new file mode 100644
index 0000000..d050e85
--- /dev/null
+++ b/source/application/api/use_case/kws/include/KwsClassifier.hpp
@@ -0,0 +1,66 @@
+/*
+ * 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_CLASSIFIER_HPP
+#define KWS_CLASSIFIER_HPP
+
+#include "ClassificationResult.hpp"
+#include "TensorFlowLiteMicro.hpp"
+#include "Classifier.hpp"
+
+#include <vector>
+
+namespace arm {
+namespace app {
+
+ /**
+ * @brief KWS Classifier - a helper class to get certain number of top
+ * results from the output vector from a classification NN.
+ * Allows for averaging of previous results.
+ **/
+ class KwsClassifier : 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. Default is 1.
+ * @param[in] useSoftmax Whether Softmax normalisation should be applied to output. Default is false.
+ * @param[in/out] resultHistory History of previous classification results to be updated.
+ * @return true if successful, false otherwise.
+ **/
+ using Classifier::GetClassificationResults; /* We are overloading not overriding. */
+ bool GetClassificationResults(TfLiteTensor* outputTensor, std::vector<ClassificationResult>& vecResults,
+ const std::vector <std::string>& labels, uint32_t topNCount,
+ bool use_softmax, std::vector<std::vector<float>>& resultHistory);
+
+ /**
+ * @brief Average the given history of results.
+ * @param[in] resultHistory The history of results to take on average of.
+ * @param[out] averageResult The calculated average.
+ **/
+ static void AveragResults(const std::vector<std::vector<float>>& resultHistory,
+ std::vector<float>& averageResult);
+ };
+
+} /* namespace app */
+} /* namespace arm */
+
+#endif /* KWS_CLASSIFIER_HPP */
diff --git a/source/application/api/use_case/kws/include/KwsProcessing.hpp b/source/application/api/use_case/kws/include/KwsProcessing.hpp
index 0ede425..e2d3ff9 100644
--- a/source/application/api/use_case/kws/include/KwsProcessing.hpp
+++ b/source/application/api/use_case/kws/include/KwsProcessing.hpp
@@ -19,7 +19,7 @@
#include "AudioUtils.hpp"
#include "BaseProcessing.hpp"
-#include "Classifier.hpp"
+#include "KwsClassifier.hpp"
#include "MicroNetKwsMfcc.hpp"
#include <functional>
@@ -55,9 +55,8 @@
* @param[in] inputSize Size of the input data.
* @return true if successful, false otherwise.
**/
- bool DoPreProcess(const void* input, size_t inputSize) override;
+ bool DoPreProcess(const void* input, size_t inferenceIndex = 0) 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. */
@@ -106,11 +105,11 @@
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. */
-
+ TfLiteTensor* m_outputTensor; /* Model output tensor. */
+ KwsClassifier& 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. */
+ std::vector<std::vector<float>> m_resultHistory; /* Store previous results so they can be averaged. */
public:
/**
* @brief Constructor
@@ -119,9 +118,9 @@
* @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,
+ KwsPostProcess(TfLiteTensor* outputTensor, KwsClassifier& classifier,
const std::vector<std::string>& labels,
- std::vector<ClassificationResult>& results);
+ std::vector<ClassificationResult>& results, size_t averagingWindowLen = 1);
/**
* @brief Should perform post-processing of the result of inference then
diff --git a/source/application/api/use_case/kws/include/KwsResult.hpp b/source/application/api/use_case/kws/include/KwsResult.hpp
index 38f32b4..e0bb868 100644
--- a/source/application/api/use_case/kws/include/KwsResult.hpp
+++ b/source/application/api/use_case/kws/include/KwsResult.hpp
@@ -1,5 +1,5 @@
/*
- * Copyright (c) 2021 Arm Limited. All rights reserved.
+ * 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");
@@ -47,7 +47,7 @@
this->m_inferenceNumber = inferenceIdx;
this->m_resultVec = ResultVec();
- for (auto & i : resultVec) {
+ for (auto& i : resultVec) {
if (i.m_normalisedVal >= this->m_threshold) {
this->m_resultVec.emplace_back(i);
}