MLECO-2599: Replace DSCNN with MicroNet for KWS
Added SoftMax function to Mathutils to allow MicroNet
to output probability as it does not nativelu have this layer.
Minor refactoring to accommodate Softmax Calculations
Extensive renaming and updating of documentation and resource download script.
Added SoftMax function to Mathutils to allow MicroNet
to output probability.
Change-Id: I7cbbda1024d14b85c9ac1beea7ca8fbffd0b6eb5
Signed-off-by: Liam Barry <liam.barry@arm.com>
diff --git a/source/application/main/Classifier.cc b/source/application/main/Classifier.cc
index c5519fb..a6ff532 100644
--- a/source/application/main/Classifier.cc
+++ b/source/application/main/Classifier.cc
@@ -24,61 +24,40 @@
#include <set>
#include <cstdint>
#include <inttypes.h>
+#include "PlatformMath.hpp"
namespace arm {
namespace app {
- template<typename T>
- void SetVectorResults(std::set<std::pair<T, uint32_t>>& topNSet,
+ void Classifier::SetVectorResults(std::set<std::pair<float, uint32_t>>& topNSet,
std::vector<ClassificationResult>& vecResults,
- TfLiteTensor* tensor,
- const std::vector <std::string>& labels) {
-
- /* For getting the floating point values, we need quantization parameters. */
- QuantParams quantParams = GetTensorQuantParams(tensor);
+ const std::vector <std::string>& labels)
+ {
/* Reset the iterator to the largest element - use reverse iterator. */
- auto topNIter = topNSet.rbegin();
- for (size_t i = 0; i < vecResults.size() && topNIter != topNSet.rend(); ++i, ++topNIter) {
- T score = topNIter->first;
- vecResults[i].m_normalisedVal = quantParams.scale * (score - quantParams.offset);
- vecResults[i].m_label = labels[topNIter->second];
- vecResults[i].m_labelIdx = topNIter->second;
- }
- }
-
- template<>
- void SetVectorResults<float>(std::set<std::pair<float, uint32_t>>& topNSet,
- std::vector<ClassificationResult>& vecResults,
- TfLiteTensor* tensor,
- const std::vector <std::string>& labels) {
- UNUSED(tensor);
- /* Reset the iterator to the largest element - use reverse iterator. */
auto topNIter = topNSet.rbegin();
for (size_t i = 0; i < vecResults.size() && topNIter != topNSet.rend(); ++i, ++topNIter) {
vecResults[i].m_normalisedVal = topNIter->first;
vecResults[i].m_label = labels[topNIter->second];
vecResults[i].m_labelIdx = topNIter->second;
}
-
}
- template<typename T>
- bool Classifier::GetTopNResults(TfLiteTensor* tensor,
+ bool Classifier::GetTopNResults(const std::vector<float>& tensor,
std::vector<ClassificationResult>& vecResults,
uint32_t topNCount,
const std::vector <std::string>& labels)
{
- std::set<std::pair<T, uint32_t>> sortedSet;
+
+ std::set<std::pair<float , uint32_t>> sortedSet;
/* NOTE: inputVec's size verification against labels should be
* checked by the calling/public function. */
- T* tensorData = tflite::GetTensorData<T>(tensor);
/* Set initial elements. */
for (uint32_t i = 0; i < topNCount; ++i) {
- sortedSet.insert({tensorData[i], i});
+ sortedSet.insert({tensor[i], i});
}
/* Initialise iterator. */
@@ -86,33 +65,26 @@
/* Scan through the rest of elements with compare operations. */
for (uint32_t i = topNCount; i < labels.size(); ++i) {
- if (setFwdIter->first < tensorData[i]) {
+ if (setFwdIter->first < tensor[i]) {
sortedSet.erase(*setFwdIter);
- sortedSet.insert({tensorData[i], i});
+ sortedSet.insert({tensor[i], i});
setFwdIter = sortedSet.begin();
}
}
/* Final results' container. */
vecResults = std::vector<ClassificationResult>(topNCount);
-
- SetVectorResults<T>(sortedSet, vecResults, tensor, labels);
+ SetVectorResults(sortedSet, vecResults, labels);
return true;
}
- template bool Classifier::GetTopNResults<uint8_t>(TfLiteTensor* tensor,
- std::vector<ClassificationResult>& vecResults,
- uint32_t topNCount, const std::vector <std::string>& labels);
-
- template bool Classifier::GetTopNResults<int8_t>(TfLiteTensor* tensor,
- std::vector<ClassificationResult>& vecResults,
- uint32_t topNCount, const std::vector <std::string>& labels);
-
bool Classifier::GetClassificationResults(
TfLiteTensor* outputTensor,
std::vector<ClassificationResult>& vecResults,
- const std::vector <std::string>& labels, uint32_t topNCount)
+ const std::vector <std::string>& labels,
+ uint32_t topNCount,
+ bool useSoftmax)
{
if (outputTensor == nullptr) {
printf_err("Output vector is null pointer.\n");
@@ -120,7 +92,7 @@
}
uint32_t totalOutputSize = 1;
- for (int inputDim = 0; inputDim < outputTensor->dims->size; inputDim++){
+ for (int inputDim = 0; inputDim < outputTensor->dims->size; inputDim++) {
totalOutputSize *= outputTensor->dims->data[inputDim];
}
@@ -139,22 +111,52 @@
bool resultState;
vecResults.clear();
- /* Get the top N results. */
+ /* De-Quantize Output Tensor */
+ QuantParams quantParams = GetTensorQuantParams(outputTensor);
+
+ /* Floating point tensor data to be populated
+ * NOTE: The assumption here is that the output tensor size isn't too
+ * big and therefore, there's neglibible impact on heap usage. */
+ std::vector<float> tensorData(totalOutputSize);
+
+ /* Populate the floating point buffer */
switch (outputTensor->type) {
- case kTfLiteUInt8:
- resultState = GetTopNResults<uint8_t>(outputTensor, vecResults, topNCount, labels);
+ case kTfLiteUInt8: {
+ uint8_t *tensor_buffer = tflite::GetTensorData<uint8_t>(outputTensor);
+ for (size_t i = 0; i < totalOutputSize; ++i) {
+ tensorData[i] = quantParams.scale *
+ (static_cast<float>(tensor_buffer[i]) - quantParams.offset);
+ }
break;
- case kTfLiteInt8:
- resultState = GetTopNResults<int8_t>(outputTensor, vecResults, topNCount, labels);
+ }
+ case kTfLiteInt8: {
+ int8_t *tensor_buffer = tflite::GetTensorData<int8_t>(outputTensor);
+ for (size_t i = 0; i < totalOutputSize; ++i) {
+ tensorData[i] = quantParams.scale *
+ (static_cast<float>(tensor_buffer[i]) - quantParams.offset);
+ }
break;
- case kTfLiteFloat32:
- resultState = GetTopNResults<float>(outputTensor, vecResults, topNCount, labels);
+ }
+ case kTfLiteFloat32: {
+ float *tensor_buffer = tflite::GetTensorData<float>(outputTensor);
+ for (size_t i = 0; i < totalOutputSize; ++i) {
+ tensorData[i] = tensor_buffer[i];
+ }
break;
+ }
default:
- printf_err("Tensor type %s not supported by classifier\n", TfLiteTypeGetName(outputTensor->type));
+ printf_err("Tensor type %s not supported by classifier\n",
+ TfLiteTypeGetName(outputTensor->type));
return false;
}
+ if (useSoftmax) {
+ math::MathUtils::SoftmaxF32(tensorData);
+ }
+
+ /* Get the top N results. */
+ resultState = GetTopNResults(tensorData, vecResults, topNCount, labels);
+
if (!resultState) {
printf_err("Failed to get top N results set\n");
return false;
@@ -162,6 +164,5 @@
return true;
}
-
} /* namespace app */
} /* namespace arm */
\ No newline at end of file
diff --git a/source/application/main/PlatformMath.cc b/source/application/main/PlatformMath.cc
index 0b8882a..26b4b72 100644
--- a/source/application/main/PlatformMath.cc
+++ b/source/application/main/PlatformMath.cc
@@ -15,6 +15,8 @@
* limitations under the License.
*/
#include "PlatformMath.hpp"
+#include <algorithm>
+#include <numeric>
#if 0 == ARM_DSP_AVAILABLE
#include <cmath>
@@ -290,6 +292,24 @@
return true;
}
+ void MathUtils::SoftmaxF32(std::vector<float>& vec)
+ {
+ /* Fix for numerical stability and apply exp. */
+ auto start = vec.begin();
+ auto end = vec.end();
+
+ float maxValue = *std::max_element(start, end);
+ for (auto it = start; it != end; ++it) {
+ *it = std::exp((*it) - maxValue);
+ }
+
+ float sumExp = std::accumulate(start, end, 0.0f);
+
+ for (auto it = start; it != end; ++it) {
+ *it = (*it)/sumExp;
+ }
+ }
+
} /* namespace math */
} /* namespace app */
} /* namespace arm */
diff --git a/source/application/main/include/Classifier.hpp b/source/application/main/include/Classifier.hpp
index 3ee3148..d899e8e 100644
--- a/source/application/main/include/Classifier.hpp
+++ b/source/application/main/include/Classifier.hpp
@@ -42,18 +42,33 @@
* 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.
* @return true if successful, false otherwise.
**/
+
virtual bool GetClassificationResults(
TfLiteTensor* outputTensor,
std::vector<ClassificationResult>& vecResults,
- const std::vector <std::string>& labels, uint32_t topNCount);
+ const std::vector <std::string>& labels, uint32_t topNCount,
+ bool use_softmax = false);
+
+ /**
+ * @brief Populate the elements of the Classification Result object.
+ * @param[in] topNSet Ordered set of top 5 output class scores and labels.
+ * @param[out] vecResults A vector of classification results.
+ * populated by this function.
+ * @param[in] labels Labels vector to match classified classes.
+ **/
+
+ void SetVectorResults(
+ std::set<std::pair<float, uint32_t>>& topNSet,
+ std::vector<ClassificationResult>& vecResults,
+ const std::vector <std::string>& labels);
private:
/**
* @brief Utility function that gets the top N classification results from the
* output vector.
- * @tparam T value type
* @param[in] tensor Inference output tensor from an NN model.
* @param[out] vecResults A vector of classification results
* populated by this function.
@@ -61,8 +76,8 @@
* @param[in] labels Labels vector to match classified classes.
* @return true if successful, false otherwise.
**/
- template<typename T>
- bool GetTopNResults(TfLiteTensor* tensor,
+
+ bool GetTopNResults(const std::vector<float>& tensor,
std::vector<ClassificationResult>& vecResults,
uint32_t topNCount,
const std::vector <std::string>& labels);
diff --git a/source/application/main/include/PlatformMath.hpp b/source/application/main/include/PlatformMath.hpp
index 6804025..fdb51b2 100644
--- a/source/application/main/include/PlatformMath.hpp
+++ b/source/application/main/include/PlatformMath.hpp
@@ -161,7 +161,14 @@
float* ptrDst,
const uint32_t dstLen);
+ /**
+ * @brief Scales output scores for an arbitrary number of classes so
+ * that they sum to 1, allowing output to be expressed as a probability.
+ * @param[in] vector Vector of floats modified in-place
+ */
+ static void SoftmaxF32(std::vector<float>& vec);
};
+
} /* namespace math */
} /* namespace app */
} /* namespace arm */