COMPMID-1310: Create graph validation executables.

Change-Id: I9e0b57b1b83fe5a95777cdaeddba6ecef650bafc
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/138697
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

1 files changed, 217 insertions, 119 deletions

diff --git a/examples/graph_mobilenet.cpp b/examples/graph_mobilenet.cpp
index 40243bb111..9304b2b380 100644
--- a/examples/graph_mobilenet.cpp
+++ b/examples/graph_mobilenet.cpp
@@ -23,11 +23,11 @@
  */
 #include "arm_compute/graph.h"
 #include "support/ToolchainSupport.h"
+#include "utils/CommonGraphOptions.h"
 #include "utils/GraphUtils.h"
 #include "utils/Utils.h"
 
-#include <cstdlib>
-
+using namespace arm_compute;
 using namespace arm_compute::utils;
 using namespace arm_compute::graph::frontend;
 using namespace arm_compute::graph_utils;
@@ -35,107 +35,112 @@ using namespace arm_compute::graph_utils;
 /** Example demonstrating how to implement MobileNet's network using the Compute Library's graph API
  *
  * @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner), [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160), [optional] Path to the weights folder, [optional] image, [optional] labels, [optional] data layout, [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
+ * @param[in] argv Arguments
  */
-class GraphMobilenetExample : public Example
+class GraphMobilenetExample : public Example // NOLINT
 {
 public:
-    void do_setup(int argc, char **argv) override
+    GraphMobilenetExample()
+        : cmd_parser(), common_opts(cmd_parser), common_params(), graph(0, "MobileNetV1")
     {
-        std::string data_path; /* Path to the trainable data */
-        std::string image;     /* Image data */
-        std::string label;     /* Label data */
-
-        // Create a preprocessor object
-        std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
+        // Add model id option
+        model_id_opt = cmd_parser.add_option<SimpleOption<int>>("model-id", 0);
+        model_id_opt->set_help("Mobilenet model id (0: 1.0_224, else: 0.75_160");
+    }
+    /** Prevent instances of this class from being copy constructed */
+    GraphMobilenetExample(const GraphMobilenetExample &) = delete;
+    /** Prevent instances of this class from being copied */
+    GraphMobilenetExample &operator=(const GraphMobilenetExample &) = delete;
+    bool do_setup(int argc, char **argv) override
+    {
+        // Parse arguments
+        cmd_parser.parse(argc, argv);
 
-        // Set target. 0 (NEON), 1 (OpenCL), 2 (OpenCL with Tuner). By default it is NEON
-        const int                  target                     = argc > 1 ? std::strtol(argv[1], nullptr, 10) : 0;
-        Target                     target_hint                = set_target_hint(target);
-        DepthwiseConvolutionMethod depthwise_convolution_hint = DepthwiseConvolutionMethod::OPTIMIZED_3x3;
-        FastMathHint               fast_math_hint             = FastMathHint::DISABLED;
-
-        // Set model to execute. 0 (MobileNetV1_1.0_224), 1 (MobileNetV1_0.75_160)
-        int model_id = (argc > 2) ? std::strtol(argv[2], nullptr, 10) : 0;
-        ARM_COMPUTE_ERROR_ON_MSG(model_id > 1, "Invalid model ID. Model must be 0 (MobileNetV1_1.0_224) or 1 (MobileNetV1_0.75_160)");
-        int layout_id = (argc > 3) ? std::strtol(argv[3], nullptr, 10) : 0;
-        ARM_COMPUTE_ERROR_ON_MSG(layout_id > 1, "Invalid layout ID. Layout must be 0 (NCHW) or 1 (NHWC)");
-
-        float            depth_scale           = (model_id == 0) ? 1.f : 0.75;
-        unsigned int     spatial_size          = (model_id == 0) ? 224 : 160;
-        std::string      model_path            = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
-        TensorDescriptor input_descriptor_nchw = TensorDescriptor(TensorShape(spatial_size, spatial_size, 3U, 1U), DataType::F32);
-        TensorDescriptor input_descriptor_nhwc = TensorDescriptor(TensorShape(3U, spatial_size, spatial_size, 1U), DataType::F32).set_layout(DataLayout::NHWC);
-        TensorDescriptor input_descriptor      = (layout_id == 0) ? input_descriptor_nchw : input_descriptor_nhwc;
+        // Consume common parameters
+        common_params = consume_common_graph_parameters(common_opts);
 
-        // Parse arguments
-        if(argc < 2)
-        {
-            // Print help
-            std::cout << "Usage: " << argv[0] << " [target] [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
-            std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
-            std::cout << "No data layout provided: using NCHW\n\n";
-            std::cout << "No data folder provided: using random values\n\n";
-        }
-        else if(argc == 2)
+        // Return when help menu is requested
+        if(common_params.help)
         {
-            std::cout << "Usage: " << argv[0] << " " << argv[1] << " [model] [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
-            std::cout << "No model ID provided: using MobileNetV1_1.0_224\n\n";
-            std::cout << "No data layout provided: using NCHW\n\n";
-            std::cout << "No data folder provided: using random values\n\n";
+            cmd_parser.print_help(argv[0]);
+            return false;
         }
-        else if(argc == 3)
-        {
-            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [layout] [path_to_data] [image] [labels] [fast_math_hint]\n\n";
-            std::cout << "No data layout provided: using NCHW\n\n";
-            std::cout << "No data folder provided: using random values\n\n";
-        }
-        else if(argc == 4)
-        {
-            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [path_to_data] [image] [labels] [fast_math_hint]\n\n";
-            std::cout << "No data folder provided: using random values\n\n";
-        }
-        else if(argc == 5)
-        {
-            data_path = argv[4];
-            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [image] [labels] [fast_math_hint]\n\n";
-            std::cout << "No image provided: using random values\n\n";
-            std::cout << "No text file with labels provided: skipping output accessor\n\n";
-        }
-        else if(argc == 6)
+
+        // Print parameter values
+        std::cout << common_params << std::endl;
+
+        // Get model parameters
+        int model_id = model_id_opt->value();
+
+        // Create input descriptor
+        unsigned int spatial_size = (model_id == 0 || common_params.data_type == DataType::QASYMM8) ? 224 : 160;
+        TensorShape  tensor_shape = TensorShape(spatial_size, spatial_size, 3U, 1U);
+        if(common_params.data_layout == DataLayout::NHWC)
         {
-            data_path = argv[4];
-            image     = argv[5];
-            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " [labels] [fast_math_hint]\n\n";
-            std::cout << "No text file with labels provided: skipping output accessor\n\n";
+            arm_compute::permute(tensor_shape, arm_compute::PermutationVector(2U, 0U, 1U));
         }
-        else if(argc == 7)
+        TensorDescriptor input_descriptor = TensorDescriptor(tensor_shape, common_params.data_type).set_layout(common_params.data_layout);
+
+        // Set graph hints
+        graph << common_params.target
+              << DepthwiseConvolutionMethod::OPTIMIZED_3x3 // FIXME(COMPMID-1073): Add heuristics to automatically call the optimized 3x3 method
+              << common_params.fast_math_hint;
+
+        // Create core graph
+        if(arm_compute::is_data_type_float(common_params.data_type))
         {
-            data_path = argv[4];
-            image     = argv[5];
-            label     = argv[6];
-            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " " << argv[3] << " " << argv[4] << " [fast_math_hint]\n\n";
-            std::cout << "No fast math info provided: disabling fast math\n\n";
+            create_graph_float(input_descriptor, model_id);
         }
         else
         {
-            data_path      = argv[4];
-            image          = argv[5];
-            label          = argv[6];
-            fast_math_hint = (std::strtol(argv[7], nullptr, 1) == 0) ? FastMathHint::DISABLED : FastMathHint::ENABLED;
+            create_graph_qasymm(input_descriptor);
         }
 
+        // Create common tail
+        graph << ReshapeLayer(TensorShape(1001U)).set_name("Reshape")
+              << SoftmaxLayer().set_name("Softmax")
+              << OutputLayer(get_output_accessor(common_params, 5));
+
+        // Finalize graph
+        GraphConfig config;
+        config.num_threads = common_params.threads;
+        config.use_tuner   = common_params.enable_tuner;
+        graph.finalize(common_params.target, config);
+
+        return true;
+    }
+    void do_run() override
+    {
+        // Run graph
+        graph.run();
+    }
+
+private:
+    CommandLineParser  cmd_parser;
+    CommonGraphOptions common_opts;
+    SimpleOption<int> *model_id_opt{ nullptr };
+    CommonGraphParams  common_params;
+    Stream             graph;
+
+    void create_graph_float(TensorDescriptor &input_descriptor, int model_id)
+    {
+        float       depth_scale = (model_id == 0) ? 1.f : 0.75;
+        std::string model_path  = (model_id == 0) ? "/cnn_data/mobilenet_v1_1_224_model/" : "/cnn_data/mobilenet_v1_075_160_model/";
+
+        // Create a preprocessor object
+        std::unique_ptr<IPreprocessor> preprocessor = arm_compute::support::cpp14::make_unique<TFPreproccessor>();
+
+        // Get trainable parameters data path
+        std::string data_path = common_params.data_path;
+
         // Add model path to data path
         if(!data_path.empty())
         {
             data_path += model_path;
         }
 
-        graph << target_hint
-              << depthwise_convolution_hint
-              << fast_math_hint
-              << InputLayer(input_descriptor,
-                            get_input_accessor(image, std::move(preprocessor), false))
+        graph << InputLayer(input_descriptor,
+                            get_input_accessor(common_params, std::move(preprocessor), false))
               << ConvolutionLayer(
                   3U, 3U, 32U * depth_scale,
                   get_weights_accessor(data_path, "Conv2d_0_weights.npy", DataLayout::NCHW),
@@ -150,47 +155,122 @@ public:
                   0.001f)
               .set_name("Conv2d_0/BatchNorm")
               << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f)).set_name("Conv2d_0/Relu6");
-        graph << get_dwsc_node(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
-        graph << get_dwsc_node(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_1", 64 * depth_scale, PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_2", 128 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_3", 128 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_4", 256 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_5", 256 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_6", 512 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_7", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_8", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_9", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_10", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_11", 512 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_12", 1024 * depth_scale, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
+        graph << get_dwsc_node_float(data_path, "Conv2d_13", 1024 * depth_scale, PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::CEIL), PadStrideInfo(1, 1, 0, 0));
         graph << PoolingLayer(PoolingLayerInfo(PoolingType::AVG)).set_name("Logits/AvgPool_1a")
               << ConvolutionLayer(
                   1U, 1U, 1001U,
                   get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_weights.npy", DataLayout::NCHW),
                   get_weights_accessor(data_path, "Logits_Conv2d_1c_1x1_biases.npy"),
                   PadStrideInfo(1, 1, 0, 0))
-              .set_name("Logits/Conv2d_1c_1x1")
-              << ReshapeLayer(TensorShape(1001U)).set_name("Reshape")
-              << SoftmaxLayer().set_name("Softmax")
-              << OutputLayer(get_output_accessor(label, 5));
-
-        // Finalize graph
-        GraphConfig config;
-        config.use_tuner = (target == 2);
-        graph.finalize(target_hint, config);
+              .set_name("Logits/Conv2d_1c_1x1");
     }
-    void do_run() override
+
+    void create_graph_qasymm(TensorDescriptor &input_descriptor)
     {
-        // Run graph
-        graph.run();
-    }
+        // Get trainable parameters data path
+        std::string data_path = common_params.data_path;
 
-private:
-    Stream graph{ 0, "MobileNetV1" };
+        // Quantization info taken from the AndroidNN QASYMM8 MobileNet example
+        const QuantizationInfo in_quant_info  = QuantizationInfo(0.0078125f, 128);
+        const QuantizationInfo mid_quant_info = QuantizationInfo(0.0784313753247f, 128);
+
+        const std::vector<QuantizationInfo> conv_weights_quant_info =
+        {
+            QuantizationInfo(0.031778190285f, 156), // conv0
+            QuantizationInfo(0.00604454148561f, 66) // conv14
+        };
+
+        const std::vector<QuantizationInfo> depth_weights_quant_info =
+        {
+            QuantizationInfo(0.254282623529f, 129),  // dwsc1
+            QuantizationInfo(0.12828284502f, 172),   // dwsc2
+            QuantizationInfo(0.265911251307f, 83),   // dwsc3
+            QuantizationInfo(0.0985597148538f, 30),  // dwsc4
+            QuantizationInfo(0.0631204470992f, 54),  // dwsc5
+            QuantizationInfo(0.0137207424268f, 141), // dwsc6
+            QuantizationInfo(0.0817828401923f, 125), // dwsc7
+            QuantizationInfo(0.0393880493939f, 164), // dwsc8
+            QuantizationInfo(0.211694166064f, 129),  // dwsc9
+            QuantizationInfo(0.158015936613f, 103),  // dwsc10
+            QuantizationInfo(0.0182712618262f, 137), // dwsc11
+            QuantizationInfo(0.0127998134121f, 134), // dwsc12
+            QuantizationInfo(0.299285322428f, 161)   // dwsc13
+        };
 
-    BranchLayer get_dwsc_node(const std::string &data_path, std::string &&param_path,
-                              unsigned int  conv_filt,
-                              PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info)
+        const std::vector<QuantizationInfo> point_weights_quant_info =
+        {
+            QuantizationInfo(0.0425766184926f, 129),  // dwsc1
+            QuantizationInfo(0.0250773020089f, 94),   // dwsc2
+            QuantizationInfo(0.015851572156f, 93),    // dwsc3
+            QuantizationInfo(0.0167811904103f, 98),   // dwsc4
+            QuantizationInfo(0.00951790809631f, 135), // dwsc5
+            QuantizationInfo(0.00999817531556f, 128), // dwsc6
+            QuantizationInfo(0.00590536883101f, 126), // dwsc7
+            QuantizationInfo(0.00576109671965f, 133), // dwsc8
+            QuantizationInfo(0.00830461271107f, 142), // dwsc9
+            QuantizationInfo(0.0152327232063f, 72),   // dwsc10
+            QuantizationInfo(0.00741417845711f, 125), // dwsc11
+            QuantizationInfo(0.0135628981516f, 142),  // dwsc12
+            QuantizationInfo(0.0338749065995f, 140)   // dwsc13
+        };
+
+        graph << InputLayer(input_descriptor.set_quantization_info(in_quant_info),
+                            get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/" + common_params.image))
+              << ConvolutionLayer(
+                  3U, 3U, 32U,
+                  get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_weights.npy"),
+                  get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Conv2d_0_bias.npy"),
+                  PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR),
+                  1, conv_weights_quant_info.at(0), mid_quant_info)
+              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_1", 64U, PadStrideInfo(1U, 1U, 1U, 1U), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(0), point_weights_quant_info.at(0));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_2", 128U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(1),
+                                      point_weights_quant_info.at(1));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_3", 128U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(2),
+                                      point_weights_quant_info.at(2));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_4", 256U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(3),
+                                      point_weights_quant_info.at(3));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_5", 256U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(4),
+                                      point_weights_quant_info.at(4));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_6", 512U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(5),
+                                      point_weights_quant_info.at(5));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_7", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(6),
+                                      point_weights_quant_info.at(6));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_8", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(7),
+                                      point_weights_quant_info.at(7));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_9", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(8),
+                                      point_weights_quant_info.at(8));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_10", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(9),
+                                      point_weights_quant_info.at(9));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_11", 512U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(10),
+                                      point_weights_quant_info.at(10));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_12", 1024U, PadStrideInfo(2U, 2U, 0U, 1U, 0U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(11),
+                                      point_weights_quant_info.at(11));
+        graph << get_dwsc_node_qasymm(data_path, "Conv2d_13", 1024U, PadStrideInfo(1U, 1U, 1U, 1U, 1U, 1U, DimensionRoundingType::FLOOR), PadStrideInfo(1U, 1U, 0U, 0U), depth_weights_quant_info.at(12),
+                                      point_weights_quant_info.at(12))
+              << PoolingLayer(PoolingLayerInfo(PoolingType::AVG))
+              << ConvolutionLayer(
+                  1U, 1U, 1001U,
+                  get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_weights.npy"),
+                  get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/Logits_Conv2d_1c_1x1_bias.npy"),
+                  PadStrideInfo(1U, 1U, 0U, 0U), 1, conv_weights_quant_info.at(1));
+    }
+
+    BranchLayer get_dwsc_node_float(const std::string &data_path, std::string &&param_path,
+                                    unsigned int  conv_filt,
+                                    PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info)
     {
         std::string total_path = param_path + "_";
         SubStream   sg(graph);
@@ -225,18 +305,36 @@ private:
 
         return BranchLayer(std::move(sg));
     }
+
+    BranchLayer get_dwsc_node_qasymm(const std::string &data_path, std::string &&param_path,
+                                     const unsigned int conv_filt,
+                                     PadStrideInfo dwc_pad_stride_info, PadStrideInfo conv_pad_stride_info,
+                                     QuantizationInfo depth_weights_quant_info, QuantizationInfo point_weights_quant_info)
+    {
+        std::string total_path = "/cnn_data/mobilenet_qasymm8_model/" + param_path + "_";
+        SubStream   sg(graph);
+
+        sg << DepthwiseConvolutionLayer(
+               3U, 3U,
+               get_weights_accessor(data_path, total_path + "depthwise_weights.npy"),
+               get_weights_accessor(data_path, total_path + "depthwise_bias.npy"),
+               dwc_pad_stride_info, depth_weights_quant_info)
+           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f))
+           << ConvolutionLayer(
+               1U, 1U, conv_filt,
+               get_weights_accessor(data_path, total_path + "pointwise_weights.npy"),
+               get_weights_accessor(data_path, total_path + "pointwise_bias.npy"),
+               conv_pad_stride_info, 1, point_weights_quant_info)
+           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f));
+
+        return BranchLayer(std::move(sg));
+    }
 };
 
 /** Main program for MobileNetV1
  *
  * @param[in] argc Number of arguments
- * @param[in] argv Arguments ( [optional] Target (0 = NEON, 1 = OpenCL, 2 = OpenCL with Tuner),
- *                             [optional] Model ID (0 = MobileNetV1_1.0_224, 1 = MobileNetV1_0.75_160),
- *                             [optional] Path to the weights folder,
- *                             [optional] image,
- *                             [optional] labels,
- *                             [optional] data layout,
- *                             [optional] Fast math for convolution layer (0 = DISABLED, 1 = ENABLED) )
+ * @param[in] argv Arguments
  */
 int main(int argc, char **argv)
 {