mv_inference: Add quantized model inference support

This patch adds quantized model inference support.
In case of quantized model, output tensor buffer contains uint8 values
between 0 and 255 so we need to convert the values to floating point ones
between 0 and 1.

Change-Id: I814b236370608d29c9b0f82f298d221851721921
Signed-off-by: Inki Dae <inki.dae@samsung.com>

diff --git a/mv_inference/inference/src/Inference.cpp b/mv_inference/inference/src/Inference.cpp
index 08600167039519f02dbd7bdeea4591cf026394dd..15d95a5db6007b431e8bc9c2a59f8a2c4e119c56 100755 (executable)
--- a/mv_inference/inference/src/Inference.cpp
+++ b/mv_inference/inference/src/Inference.cpp
@@ -758,7 +758,22 @@ int Inference::GetClassficationResults(ImageClassificationResults *classificatio
                }
 
                outputData.dimInfo.push_back(tmpDimInfo);
-               outputData.data.push_back((void *)mOutputTensorBuffers.front().buffer);
+
+               // Normalize output tensor data converting it to float type in case of quantized model.
+               if (tensor_info.data_type == TENSOR_DATA_TYPE_UINT8) {
+                       unsigned char *ori_buf = (unsigned char *)mOutputTensorBuffers[i].buffer;
+                       float *new_buf = new float[tensor_info.size];
+
+                       for (int j = 0; j < tensor_info.size; j++) {
+                               new_buf[j] = (float)ori_buf[j] / 255.0f;
+                       }
+
+                       // replace original buffer with new one, and release origin one.
+                       mOutputTensorBuffers[i].buffer = new_buf;
+                       delete[] ori_buf;
+               }
+
+               outputData.data.push_back((void *)mOutputTensorBuffers[i].buffer);
        }
 
        // Will contain top N results in ascending order.
@@ -766,17 +781,18 @@ int Inference::GetClassficationResults(ImageClassificationResults *classificatio
        std::priority_queue<std::pair<float, int>,
                std::vector<std::pair<float, int>>,
                std::greater<std::pair<float, int>>> top_result_pq;
-       float value;
+       float value = 0.0f;
 
        std::vector<std::vector<int>> inferDimInfo(outputData.dimInfo);
        std::vector<void*> inferResults(outputData.data.begin(), outputData.data.end());
 
-       long count = inferDimInfo[0][1];
-       LOGI("count: %ld", count);
+       int count = inferDimInfo[0][1];
+       LOGI("count: %d", count);
 
        float *prediction = reinterpret_cast<float*>(inferResults[0]);
        for (int i = 0; i < count; ++i) {
                value = prediction[i];
+
                // Only add it if it beats the threshold and has a chance at being in
                // the top N.
                top_result_pq.push(std::pair<float, int>(value, i));