COMPMID-1849: Implement CPPDetectionPostProcessLayer

* Add DetectionPostProcessLayer
* Add DetectionPostProcessLayer at the graph

Change-Id: I7e56f6cffc26f112d26dfe74853085bb8ec7d849
Signed-off-by: Isabella Gottardi <isabella.gottardi@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1639
Reviewed-by: Giuseppe Rossini <giuseppe.rossini@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 63 insertions, 29 deletions

diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp
index dad9aed6a5..00165cd6c2 100644
--- a/utils/GraphUtils.cpp
+++ b/utils/GraphUtils.cpp
@@ -140,12 +140,14 @@ bool DummyAccessor::access_tensor(ITensor &tensor)
     return ret;
 }
 
-NumPyAccessor::NumPyAccessor(std::string npy_path, TensorShape shape, DataType data_type, std::ostream &output_stream)
+NumPyAccessor::NumPyAccessor(std::string npy_path, TensorShape shape, DataType data_type, DataLayout data_layout, std::ostream &output_stream)
     : _npy_tensor(), _filename(std::move(npy_path)), _output_stream(output_stream)
 {
-    NumPyBinLoader loader(_filename);
+    NumPyBinLoader loader(_filename, data_layout);
 
     TensorInfo info(shape, 1, data_type);
+    info.set_data_layout(data_layout);
+
     _npy_tensor.allocator()->init(info);
     _npy_tensor.allocator()->allocate();
 
diff --git a/utils/GraphUtils.h b/utils/GraphUtils.h
index fe19eb3196..3417135f17 100644
--- a/utils/GraphUtils.h
+++ b/utils/GraphUtils.h
@@ -145,9 +145,10 @@ public:
      * @param[in]  npy_path      Path to npy file.
      * @param[in]  shape         Shape of the numpy tensor data.
      * @param[in]  data_type     DataType of the numpy tensor data.
+     * @param[in]  data_layout   (Optional) DataLayout of the numpy tensor data.
      * @param[out] output_stream (Optional) Output stream
      */
-    NumPyAccessor(std::string npy_path, TensorShape shape, DataType data_type, std::ostream &output_stream = std::cout);
+    NumPyAccessor(std::string npy_path, TensorShape shape, DataType data_type, DataLayout data_layout = DataLayout::NCHW, std::ostream &output_stream = std::cout);
     /** Allow instances of this class to be move constructed */
     NumPyAccessor(NumPyAccessor &&) = default;
     /** Prevent instances of this class from being copied (As this class contains pointers) */
@@ -567,11 +568,13 @@ inline std::unique_ptr<graph::ITensorAccessor> get_detection_output_accessor(con
  * @param[in]  npy_path      Path to npy file.
  * @param[in]  shape         Shape of the numpy tensor data.
  * @param[in]  data_type     DataType of the numpy tensor data.
+ * @param[in]  data_layout   DataLayout of the numpy tensor data.
  * @param[out] output_stream (Optional) Output stream
  *
  * @return An appropriate tensor accessor
  */
-inline std::unique_ptr<graph::ITensorAccessor> get_npy_output_accessor(const std::string &npy_path, TensorShape shape, DataType data_type, std::ostream &output_stream = std::cout)
+inline std::unique_ptr<graph::ITensorAccessor> get_npy_output_accessor(const std::string &npy_path, TensorShape shape, DataType data_type, DataLayout data_layout = DataLayout::NCHW,
+                                                                       std::ostream &output_stream = std::cout)
 {
     if(npy_path.empty())
     {
@@ -579,7 +582,7 @@ inline std::unique_ptr<graph::ITensorAccessor> get_npy_output_accessor(const std
     }
     else
     {
-        return arm_compute::support::cpp14::make_unique<NumPyAccessor>(npy_path, shape, data_type, output_stream);
+        return arm_compute::support::cpp14::make_unique<NumPyAccessor>(npy_path, shape, data_type, data_layout, output_stream);
     }
 }
 
diff --git a/utils/TypePrinter.h b/utils/TypePrinter.h
index 6ba6f45f8c..f51d2368e1 100644
--- a/utils/TypePrinter.h
+++ b/utils/TypePrinter.h
@@ -2059,6 +2059,43 @@ inline std::string to_string(const DetectionOutputLayerInfo &detection_info)
     str << detection_info;
     return str.str();
 }
+/** Formatted output of the DetectionPostProcessLayerInfo type.
+ *
+ * @param[out] os             Output stream
+ * @param[in]  detection_info Type to output
+ *
+ * @return Modified output stream.
+ */
+inline ::std::ostream &operator<<(::std::ostream &os, const DetectionPostProcessLayerInfo &detection_info)
+{
+    os << "{MaxDetections=" << detection_info.max_detections() << ","
+       << "MaxClassesPerDetection=" << detection_info.max_classes_per_detection() << ","
+       << "NmsScoreThreshold=" << detection_info.nms_score_threshold() << ","
+       << "NmsIouThreshold=" << detection_info.iou_threshold() << ","
+       << "NumClasses=" << detection_info.num_classes() << ","
+       << "ScaleValue_y=" << detection_info.scale_value_y() << ","
+       << "ScaleValue_x=" << detection_info.scale_value_x() << ","
+       << "ScaleValue_h=" << detection_info.scale_value_h() << ","
+       << "ScaleValue_w=" << detection_info.scale_value_w() << ","
+       << "UseRegularNms=" << detection_info.use_regular_nms() << ","
+       << "DetectionPerClass=" << detection_info.detection_per_class()
+       << "}";
+
+    return os;
+}
+
+/** Formatted output of the DetectionPostProcessLayerInfo type.
+ *
+ * @param[in] detection_info Type to output
+ *
+ * @return Formatted string.
+ */
+inline std::string to_string(const DetectionPostProcessLayerInfo &detection_info)
+{
+    std::stringstream str;
+    str << detection_info;
+    return str.str();
+}
 /** Formatted output of the DetectionWindow type.
  *
  * @param[in] detection_window Type to output
diff --git a/utils/Utils.h b/utils/Utils.h
index ba10d7c803..cc5dfbabc2 100644
--- a/utils/Utils.h
+++ b/utils/Utils.h
@@ -616,10 +616,10 @@ void save_to_ppm(T &tensor, const std::string &ppm_filename)
  * @param[in] npy_filename  Filename of the file to create.
  * @param[in] fortran_order If true, save matrix in fortran order.
  */
-template <typename T>
+template <typename T, typename U = float>
 void save_to_npy(T &tensor, const std::string &npy_filename, bool fortran_order)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_NOT_IN(&tensor, arm_compute::DataType::F32);
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_NOT_IN(&tensor, arm_compute::DataType::F32, arm_compute::DataType::QASYMM8);
 
     std::ofstream fs;
     try
@@ -637,33 +637,25 @@ void save_to_npy(T &tensor, const std::string &npy_filename, bool fortran_order)
         // Map buffer if creating a CLTensor
         map(tensor, true);
 
-        switch(tensor.info()->data_type())
-        {
-            case arm_compute::DataType::F32:
-            {
-                std::vector<float> tmp; /* Used only to get the typestring */
-                npy::Typestring    typestring_o{ tmp };
-                std::string        typestring = typestring_o.str();
+        using typestring_type = typename std::conditional<std::is_floating_point<U>::value, float, qasymm8_t>::type;
 
-                std::ofstream stream(npy_filename, std::ofstream::binary);
-                npy::write_header(stream, typestring, fortran_order, shape);
+        std::vector<typestring_type> tmp; /* Used only to get the typestring */
+        npy::Typestring              typestring_o{ tmp };
+        std::string                  typestring = typestring_o.str();
 
-                arm_compute::Window window;
-                window.use_tensor_dimensions(tensor.info()->tensor_shape());
+        std::ofstream stream(npy_filename, std::ofstream::binary);
+        npy::write_header(stream, typestring, fortran_order, shape);
 
-                arm_compute::Iterator in(&tensor, window);
+        arm_compute::Window window;
+        window.use_tensor_dimensions(tensor.info()->tensor_shape());
 
-                arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id)
-                {
-                    stream.write(reinterpret_cast<const char *>(in.ptr()), sizeof(float));
-                },
-                in);
+        arm_compute::Iterator in(&tensor, window);
 
-                break;
-            }
-            default:
-                ARM_COMPUTE_ERROR("Unsupported format");
-        }
+        arm_compute::execute_window_loop(window, [&](const arm_compute::Coordinates & id)
+        {
+            stream.write(reinterpret_cast<const char *>(in.ptr()), sizeof(typestring_type));
+        },
+        in);
 
         // Unmap buffer if creating a CLTensor
         unmap(tensor);