IVGCVSW-2560 Verify Inference test for TensorFlow Lite MobileNet SSD

 * Assign output shape of MobileNet SSD to ArmNN network
 * Add m_OverridenOutputShapes to TfLiteParser to set shape in GetNetworkOutputBindingInfo
 * Use input quantization instead of output quantization params
 * Correct data and datatype in Inference test

Change-Id: I01ac2e07ed08e8928ba0df33a4847399e1dd8394
Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>

4 files changed, 28 insertions, 14 deletions

diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp
index cd0e9214c2..31e808fd6e 100644
--- a/src/armnnTfLiteParser/TfLiteParser.cpp
+++ b/src/armnnTfLiteParser/TfLiteParser.cpp
@@ -295,7 +295,7 @@ void CalcPadding(uint32_t inputSize,
     }
 }
 
-armnn::TensorInfo ToTensorInfo(TfLiteParser::TensorRawPtr tensorPtr)
+armnn::TensorInfo ToTensorInfo(TfLiteParser::TensorRawPtr tensorPtr, const std::vector<unsigned int>& shapes)
 {
     armnn::DataType type;
     CHECK_TENSOR_PTR(tensorPtr);
@@ -345,17 +345,21 @@ armnn::TensorInfo ToTensorInfo(TfLiteParser::TensorRawPtr tensorPtr)
         }
     }
 
-    auto const & dimensions = AsUnsignedVector(tensorPtr->shape);
-
     // two statements (on purpose) for easier debugging:
-    armnn::TensorInfo result(static_cast<unsigned int>(tensorPtr->shape.size()),
-                             dimensions.data(),
+    armnn::TensorInfo result(static_cast<unsigned int>(shapes.size()),
+                             shapes.data(),
                              type,
                              quantizationScale,
                              quantizationOffset);
     return result;
 }
 
+armnn::TensorInfo ToTensorInfo(TfLiteParser::TensorRawPtr tensorPtr)
+{
+    auto const & dimensions = AsUnsignedVector(tensorPtr->shape);
+    return ToTensorInfo(tensorPtr, dimensions);
+}
+
 template<typename T>
 std::pair<armnn::ConstTensor, std::unique_ptr<T[]>>
 CreateConstTensorImpl(TfLiteParser::BufferRawPtr bufferPtr,
@@ -1796,10 +1800,17 @@ void TfLiteParser::ParseDetectionPostProcess(size_t subgraphIndex, size_t operat
 
     BOOST_ASSERT(layer != nullptr);
 
-    // Register outputs
+    // The model does not specify the output shapes.
+    // The output shapes are calculated from the max_detection and max_classes_per_detection.
+    unsigned int numDetectedBox = desc.m_MaxDetections * desc.m_MaxClassesPerDetection;
+    m_OverridenOutputShapes.push_back({ 1, numDetectedBox, 4 });
+    m_OverridenOutputShapes.push_back({ 1, numDetectedBox });
+    m_OverridenOutputShapes.push_back({ 1, numDetectedBox });
+    m_OverridenOutputShapes.push_back({ 1 });
+
     for (unsigned int i = 0 ; i < outputs.size() ; ++i)
     {
-        armnn::TensorInfo detectionBoxOutputTensorInfo = ToTensorInfo(outputs[i]);
+        armnn::TensorInfo detectionBoxOutputTensorInfo = ToTensorInfo(outputs[i], m_OverridenOutputShapes[i]);
         layer->GetOutputSlot(i).SetTensorInfo(detectionBoxOutputTensorInfo);
     }
 
@@ -2232,12 +2243,15 @@ BindingPointInfo TfLiteParser::GetNetworkOutputBindingInfo(size_t subgraphId,
 {
     CHECK_SUBGRAPH(m_Model, subgraphId);
     auto outputs = GetSubgraphOutputs(m_Model, subgraphId);
-    for (auto const & output : outputs)
+    for (unsigned int i = 0; i < outputs.size(); ++i)
     {
+        auto const output = outputs[i];
         if (output.second->name == name)
         {
             auto bindingId = GenerateLayerBindingId(subgraphId, output.first);
-            return std::make_pair(bindingId, ToTensorInfo(output.second));
+            std::vector<unsigned int> shape = m_OverridenOutputShapes.size() > 0 ?
+                                                m_OverridenOutputShapes[i] : AsUnsignedVector(output.second->shape);
+            return std::make_pair(bindingId, ToTensorInfo(output.second, shape));
         }
     }
 
diff --git a/src/armnnTfLiteParser/TfLiteParser.hpp b/src/armnnTfLiteParser/TfLiteParser.hpp
index 3fe4809aa2..2895487214 100644
--- a/src/armnnTfLiteParser/TfLiteParser.hpp
+++ b/src/armnnTfLiteParser/TfLiteParser.hpp
@@ -193,6 +193,10 @@ private:
     /// Connections for tensors in each subgraph
     /// The first index is the subgraph ID, the second index is the tensor ID
     std::vector<TensorConnections> m_SubgraphConnections;
+
+    /// This is used in case that the model does not speciry the output.
+    /// The shape can be calculated from the options.
+    std::vector<std::vector<unsigned int>> m_OverridenOutputShapes;
 };
 
 }
diff --git a/src/armnnTfLiteParser/test/DetectionPostProcess.cpp b/src/armnnTfLiteParser/test/DetectionPostProcess.cpp
index 3c602937ce..3002885016 100644
--- a/src/armnnTfLiteParser/test/DetectionPostProcess.cpp
+++ b/src/armnnTfLiteParser/test/DetectionPostProcess.cpp
@@ -76,28 +76,24 @@ struct DetectionPostProcessFixture : ParserFlatbuffersFixture
                             }
                         },
                         {
-                            "shape": [1, 3, 4],
                             "type": "FLOAT32",
                             "buffer": 3,
                             "name": "detection_boxes",
                             "quantization": {}
                         },
                         {
-                            "shape": [1, 3],
                             "type": "FLOAT32",
                             "buffer": 4,
                             "name": "detection_classes",
                             "quantization": {}
                         },
                         {
-                            "shape": [1, 3],
                             "type": "FLOAT32",
                             "buffer": 5,
                             "name": "detection_scores",
                             "quantization": {}
                         },
                         {
-                            "shape": [1],
                             "type": "FLOAT32",
                             "buffer": 6,
                             "name": "num_detections",
diff --git a/src/backends/reference/workloads/DetectionPostProcess.cpp b/src/backends/reference/workloads/DetectionPostProcess.cpp
index 2eb35f5ffa..6868180b0b 100644
--- a/src/backends/reference/workloads/DetectionPostProcess.cpp
+++ b/src/backends/reference/workloads/DetectionPostProcess.cpp
@@ -197,7 +197,7 @@ void DetectionPostProcess(const TensorInfo& boxEncodingsInfo,
             }
             std::vector<unsigned int> selectedIndices = NonMaxSuppression(numBoxes, boxCorners, classScores,
                                                                           desc.m_NmsScoreThreshold,
-                                                                          desc.m_MaxClassesPerDetection,
+                                                                          desc.m_DetectionsPerClass,
                                                                           desc.m_NmsIouThreshold);
 
             for (unsigned int i = 0; i < selectedIndices.size(); ++i)