COMPMID-752 Creating an example for QASYMM8 MobileNet

Change-Id: Ic76b3b6adaff8c84ba4d2ca5283d9291c69344f0
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/114466
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Pablo Tello <pablo.tello@arm.com>
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>

10 files changed, 318 insertions, 56 deletions

diff --git a/arm_compute/graph/nodes/ConvolutionLayer.h b/arm_compute/graph/nodes/ConvolutionLayer.h
index 0905524de8..1806190971 100644
--- a/arm_compute/graph/nodes/ConvolutionLayer.h
+++ b/arm_compute/graph/nodes/ConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,24 +44,28 @@ class ConvolutionLayer final : public INode
 public:
     /** Default Constructor
      *
-     * @param[in] conv_width   Convolution width
-     * @param[in] conv_height  Convolution height
-     * @param[in] ofm          Output feature map
-     * @param[in] weights      Weights of the convolution layer
-     * @param[in] biases       Bias of the convolution layer
-     * @param[in] conv_info    Convolution information
-     * @param[in] num_groups   (Optional) Number of groups, default = 1
-     * @param[in] weights_info (Optional) Weights information
+     * @param[in] conv_width         Convolution width
+     * @param[in] conv_height        Convolution height
+     * @param[in] ofm                Output feature map
+     * @param[in] weights            Weights of the convolution layer
+     * @param[in] biases             Bias of the convolution layer
+     * @param[in] conv_info          Convolution information
+     * @param[in] num_groups         (Optional) Number of groups, default = 1
+     * @param[in] weights_info       (Optional) Weights information
+     * @param[in] weights_quant_info (Optional) Weights quantization information
+     * @param[in] out_quant_info     (Optional) Output quantization info
      */
     template <typename AccessorTypeWeights, typename AccessorTypeBiases>
-    ConvolutionLayer(unsigned int          conv_width,
-                     unsigned int          conv_height,
-                     unsigned int          ofm,
+    ConvolutionLayer(unsigned int           conv_width,
+                     unsigned int           conv_height,
+                     unsigned int           ofm,
                      AccessorTypeWeights &&weights,
-                     AccessorTypeBiases &&biases,
-                     const PadStrideInfo   conv_info,
-                     unsigned int          num_groups   = 1,
-                     const WeightsInfo     weights_info = WeightsInfo())
+                     AccessorTypeBiases   &&biases,
+                     const PadStrideInfo    conv_info,
+                     unsigned int           num_groups         = 1,
+                     const WeightsInfo      weights_info       = WeightsInfo(),
+                     const QuantizationInfo weights_quant_info = QuantizationInfo(),
+                     const QuantizationInfo out_quant_info     = QuantizationInfo())
         : _conv_width(conv_width),
           _conv_height(conv_height),
           _ofm(ofm),
@@ -70,6 +74,8 @@ public:
           _conv_info(std::move(conv_info)),
           _num_groups(num_groups),
           _weights_info(std::move(weights_info)),
+          _weights_quant_info(std::move(weights_quant_info)),
+          _out_quant_info(std::move(out_quant_info)),
           _is(nullptr),
           _os(nullptr),
           _ws(nullptr),
@@ -101,14 +107,16 @@ private:
     std::unique_ptr<arm_compute::IFunction> instantiate_grouped_convolution(ITensor *input, ITensor *output, ConvolutionMethodHint conv_method_hint);
 
 private:
-    unsigned int        _conv_width;   /**< Convolution width */
-    unsigned int        _conv_height;  /**< Convolution height */
-    unsigned int        _ofm;          /**< Output feature maps */
-    Tensor              _weights;      /**< Weights tensor */
-    Tensor              _biases;       /**< Biases tensor */
-    const PadStrideInfo _conv_info;    /**< Convolution layer information */
-    unsigned int        _num_groups;   /**< Number of groups */
-    const WeightsInfo   _weights_info; /**< Convolution layer weights information */
+    unsigned int           _conv_width;         /**< Convolution width */
+    unsigned int           _conv_height;        /**< Convolution height */
+    unsigned int           _ofm;                /**< Output feature maps */
+    Tensor                 _weights;            /**< Weights tensor */
+    Tensor                 _biases;             /**< Biases tensor */
+    const PadStrideInfo    _conv_info;          /**< Convolution layer information */
+    unsigned int           _num_groups;         /**< Number of groups */
+    const WeightsInfo      _weights_info;       /**< Convolution layer weights information */
+    const QuantizationInfo _weights_quant_info; /**< Output quantization information */
+    const QuantizationInfo _out_quant_info;     /**< Output quantization information */
 
     std::unique_ptr<SubTensor[]> _is; /**< Input tensor sub-tensors used for grouped convolution */
     std::unique_ptr<SubTensor[]> _os; /**< Output tensor sub-tensors used for grouped convolution */
diff --git a/arm_compute/graph/nodes/DepthwiseConvolutionLayer.h b/arm_compute/graph/nodes/DepthwiseConvolutionLayer.h
index 8b7e3b8296..2d4bd1e0c2 100644
--- a/arm_compute/graph/nodes/DepthwiseConvolutionLayer.h
+++ b/arm_compute/graph/nodes/DepthwiseConvolutionLayer.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -50,10 +50,12 @@ public:
      * @param[in] biases      Biases values tensor
      * @param[in] conv_info   Convolution info
      * @param[in] opt3x3      (Optional) If true executes DepthwiseConvolutionLayer3x3
+     * @param[in] quant_info  (Optional) Quantization info used for weights
      */
     template <typename AccessorType>
-    DepthwiseConvolutionLayer(unsigned int conv_width, unsigned int conv_height, AccessorType &&weights, AccessorType &&biases, const PadStrideInfo conv_info, bool opt3x3 = true)
-        : _conv_width(conv_width), _conv_height(conv_height), _weights(std::move(weights)), _biases(std::move(biases)), _conv_info(conv_info), _opt3x3(opt3x3)
+    DepthwiseConvolutionLayer(unsigned int conv_width, unsigned int conv_height, AccessorType &&weights, AccessorType &&biases, const PadStrideInfo conv_info, bool opt3x3 = true,
+                              const QuantizationInfo quant_info = QuantizationInfo())
+        : _conv_width(conv_width), _conv_height(conv_height), _weights(std::move(weights)), _biases(std::move(biases)), _conv_info(conv_info), _opt3x3(opt3x3), _quant_info(std::move(quant_info))
     {
     }
 
@@ -61,12 +63,13 @@ public:
     std::unique_ptr<arm_compute::IFunction> instantiate_node(GraphContext &ctx, ITensorObject *input, ITensorObject *output) override;
 
 private:
-    unsigned int        _conv_width;
-    unsigned int        _conv_height;
-    Tensor              _weights;
-    Tensor              _biases;
-    const PadStrideInfo _conv_info;
-    bool                _opt3x3;
+    unsigned int           _conv_width;
+    unsigned int           _conv_height;
+    Tensor                 _weights;
+    Tensor                 _biases;
+    const PadStrideInfo    _conv_info;
+    bool                   _opt3x3;
+    const QuantizationInfo _quant_info;
 };
 } // namespace graph
 } // namespace arm_compute
diff --git a/examples/graph_cl_mobilenet_qasymm8.cpp b/examples/graph_cl_mobilenet_qasymm8.cpp
new file mode 100644
index 0000000000..046c7779b1
--- /dev/null
+++ b/examples/graph_cl_mobilenet_qasymm8.cpp
@@ -0,0 +1,223 @@
+/*
+ * Copyright (c) 2017-2018 ARM Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/graph/Graph.h"
+#include "arm_compute/graph/Nodes.h"
+#include "support/ToolchainSupport.h"
+#include "utils/GraphUtils.h"
+#include "utils/Utils.h"
+
+using namespace arm_compute;
+using namespace arm_compute::graph;
+using namespace arm_compute::graph_utils;
+
+/** Example demonstrating how to implement QASYMM8 MobileNet's network using the Compute Library's graph API
+ *
+ * @param[in] argc Number of arguments
+ * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] npy_input, [optional] labels )
+ */
+class GraphMobileNetQASYMM8Example : public utils::Example
+{
+public:
+    void do_setup(int argc, char **argv) override
+    {
+        std::string data_path; /* Path to the trainable data */
+        std::string input;     /* Image data */
+        std::string label;     /* Label data */
+
+        // 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
+        };
+
+        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
+        };
+
+        // Parse arguments
+        if(argc < 2)
+        {
+            // Print help
+            std::cout << "Usage: " << argv[0] << " [path_to_data] [npy_input] [labels]\n\n";
+            std::cout << "No data folder provided: using random values\n\n";
+        }
+        else if(argc == 2)
+        {
+            data_path = argv[1];
+            std::cout << "Usage: " << argv[0] << " " << argv[1] << " [npy_input] [labels]\n\n";
+            std::cout << "No input provided: using random values\n\n";
+        }
+        else if(argc == 3)
+        {
+            data_path = argv[1];
+            input     = argv[2];
+            std::cout << "Usage: " << argv[0] << " " << argv[1] << " " << argv[2] << " [labels]\n\n";
+            std::cout << "No text file with labels provided: skipping output accessor\n\n";
+        }
+        else
+        {
+            data_path = argv[1];
+            input     = argv[2];
+            label     = argv[3];
+        }
+
+        graph << TargetHint::OPENCL
+              << arm_compute::graph::Tensor(TensorInfo(TensorShape(224U, 224U, 3U, 1U), 1, DataType::QASYMM8, in_quant_info),
+                                            get_weights_accessor(data_path, "/cnn_data/mobilenet_qasymm8_model/" + input))
+              << 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, WeightsInfo(),
+                  conv_weights_quant_info.at(0),
+                  mid_quant_info)
+              << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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))
+              << get_dwsc_node(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, WeightsInfo(), conv_weights_quant_info.at(1))
+              << ReshapeLayer(TensorShape(1001U))
+              << SoftmaxLayer()
+              << arm_compute::graph::Tensor(get_output_accessor(label, 5));
+    }
+    void do_run() override
+    {
+        // Run graph
+        graph.run();
+    }
+
+private:
+    Graph graph{};
+
+    /** This function produces a depthwise separable convolution node (i.e. depthwise + pointwise layers) with ReLU6 activation after each layer.
+     *
+     * @param[in] data_path                Path to trainable data folder
+     * @param[in] param_path               Prefix of specific set of weights/biases data
+     * @param[in] conv_filt                Filters depths for pointwise convolution
+     * @param[in] dwc_pad_stride_info      PadStrideInfo for depthwise convolution
+     * @param[in] conv_pad_stride_info     PadStrideInfo for pointwise convolution
+     * @param[in] depth_weights_quant_info QuantizationInfo for depthwise convolution's weights
+     * @param[in] point_weights_quant_info QuantizationInfo for pointwise convolution's weights
+     *
+     * @return The complete dwsc node
+     */
+    BranchLayer get_dwsc_node(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 + "_";
+        SubGraph    sg;
+
+        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,
+               true,
+               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, WeightsInfo(),
+               point_weights_quant_info)
+           << ActivationLayer(ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f));
+
+        return BranchLayer(std::move(sg));
+    }
+};
+/** Main program for MobileNetQASYMM8
+ *
+ * @param[in] argc Number of arguments
+ * @param[in] argv Arguments ( [optional] Path to the weights folder, [optional] npy_input, [optional] labels )
+ */
+int main(int argc, char **argv)
+{
+    return utils::run_example<GraphMobileNetQASYMM8Example>(argc, argv);
+}
diff --git a/src/graph/nodes/BranchLayer.cpp b/src/graph/nodes/BranchLayer.cpp
index eea0540741..6352bfc1e3 100644
--- a/src/graph/nodes/BranchLayer.cpp
+++ b/src/graph/nodes/BranchLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -53,7 +53,7 @@ void depth_concatenate_output_info(ITensorInfo *info, ITensorInfo *sub_tensor_in
         arm_compute::auto_init_if_empty(*info,
                                         sub_tensor_info->tensor_shape(),
                                         sub_tensor_info->num_channels(),
-                                        sub_tensor_info->data_type(), sub_tensor_info->fixed_point_position());
+                                        sub_tensor_info->data_type(), sub_tensor_info->fixed_point_position(), sub_tensor_info->quantization_info());
         info->set_valid_region(sub_tensor_info->valid_region());
     }
     else
@@ -170,7 +170,8 @@ std::unique_ptr<arm_compute::IFunction> BranchLayer::instantiate_node(GraphConte
                                     out_info.tensor_shape(),
                                     out_info.num_channels(),
                                     out_info.data_type(),
-                                    out_info.fixed_point_position());
+                                    out_info.fixed_point_position(),
+                                    out_info.quantization_info());
 
     return std::move(func);
 }
\ No newline at end of file
diff --git a/src/graph/nodes/ConvolutionLayer.cpp b/src/graph/nodes/ConvolutionLayer.cpp
index ae4a8d7e6b..53d06ea75f 100644
--- a/src/graph/nodes/ConvolutionLayer.cpp
+++ b/src/graph/nodes/ConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -184,14 +184,17 @@ std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_node(Graph
     // Set weights and biases info
     if(_weights.tensor() == nullptr)
     {
-        _weights.set_info(TensorInfo(TensorShape(_conv_width, _conv_height, in->info()->dimension(2) / _num_groups, _ofm),
+        TensorInfo info = TensorInfo(TensorShape(_conv_width, _conv_height, in->info()->dimension(2) / _num_groups, _ofm),
                                      in->info()->num_channels(),
                                      in->info()->data_type(),
-                                     in->info()->fixed_point_position()));
+                                     in->info()->fixed_point_position());
+        info.set_quantization_info(_weights_quant_info);
+        _weights.set_info(std::move(info));
     }
     if(_biases.has_accessor() && _biases.tensor() == nullptr)
     {
-        _biases.set_info(TensorInfo(TensorShape(_ofm), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+        DataType dt = in->info()->data_type();
+        _biases.set_info(TensorInfo(TensorShape(_ofm), in->info()->num_channels(), is_data_type_quantized_asymmetric(dt) ? DataType::S32 : dt, in->info()->fixed_point_position()));
     }
 
     std::unique_ptr<arm_compute::IFunction> func;
@@ -213,7 +216,8 @@ std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_node(Graph
     TensorShape output_shape = calculate_convolution_layer_output_shape(in->info()->tensor_shape(), _weights.info().tensor_shape(), _conv_info);
 
     // Output auto inizialitation if not yet initialized
-    arm_compute::auto_init_if_empty(*out->info(), output_shape, 1, in->info()->data_type(), in->info()->fixed_point_position());
+    arm_compute::auto_init_if_empty(*out->info(), output_shape, 1, in->info()->data_type(), in->info()->fixed_point_position(),
+                                    (_out_quant_info.empty()) ? in->info()->quantization_info() : _out_quant_info);
 
     // Create appropriate convolution function
     if(_num_groups == 1)
diff --git a/src/graph/nodes/DepthwiseConvolutionLayer.cpp b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
index b4598538c1..1209d0376e 100644
--- a/src/graph/nodes/DepthwiseConvolutionLayer.cpp
+++ b/src/graph/nodes/DepthwiseConvolutionLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -43,11 +43,14 @@ std::unique_ptr<arm_compute::IFunction> DepthwiseConvolutionLayer::instantiate_n
         TensorShape shape = in->info()->tensor_shape();
         shape.set(Window::DimX, _conv_width);
         shape.set(Window::DimY, _conv_height);
-        _weights.set_info(TensorInfo(TensorShape(shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+        TensorInfo info = TensorInfo(TensorShape(shape), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position());
+        info.set_quantization_info(_quant_info);
+        _weights.set_info(std::move(info));
     }
     if(_biases.has_accessor() && _biases.tensor() == nullptr)
     {
-        _biases.set_info(TensorInfo(TensorShape(in->info()->dimension(2)), in->info()->num_channels(), in->info()->data_type(), in->info()->fixed_point_position()));
+        DataType dt = in->info()->data_type();
+        _biases.set_info(TensorInfo(TensorShape(in->info()->dimension(2)), in->info()->num_channels(), is_data_type_quantized_asymmetric(dt) ? DataType::S32 : dt, in->info()->fixed_point_position()));
     }
 
     bool weights_is_loaded = _weights.tensor() != nullptr;
diff --git a/src/graph/nodes/ReshapeLayer.cpp b/src/graph/nodes/ReshapeLayer.cpp
index bbe0739e64..b0c117e418 100644
--- a/src/graph/nodes/ReshapeLayer.cpp
+++ b/src/graph/nodes/ReshapeLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -44,7 +44,7 @@ std::unique_ptr<arm_compute::IFunction> ReshapeLayer::instantiate_node(GraphCont
     arm_compute::ITensor *out = output->tensor();
 
     // Auto configure output
-    arm_compute::auto_init_if_empty(*out->info(), _shape, 1, in->info()->data_type(), in->info()->fixed_point_position());
+    arm_compute::auto_init_if_empty(*out->info(), _shape, 1, in->info()->data_type(), in->info()->fixed_point_position(), in->info()->quantization_info());
 
     // Create node context
     NodeContext node_ctx(OperationType::ReshapeLayer);
diff --git a/utils/GraphUtils.cpp b/utils/GraphUtils.cpp
index b9be9d4085..6b3dffc1a4 100644
--- a/utils/GraphUtils.cpp
+++ b/utils/GraphUtils.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -140,16 +140,14 @@ TopNPredictionsAccessor::TopNPredictionsAccessor(const std::string &labels_path,
     }
 }
 
-bool TopNPredictionsAccessor::access_tensor(ITensor &tensor)
+template <typename T>
+void TopNPredictionsAccessor::access_predictions_tensor(ITensor &tensor)
 {
-    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::F32);
-    ARM_COMPUTE_ERROR_ON(_labels.size() != tensor.info()->dimension(0));
-
     // Get the predicted class
-    std::vector<float>  classes_prob;
+    std::vector<T>      classes_prob;
     std::vector<size_t> index;
 
-    const auto   output_net  = reinterpret_cast<float *>(tensor.buffer() + tensor.info()->offset_first_element_in_bytes());
+    const auto   output_net  = reinterpret_cast<T *>(tensor.buffer() + tensor.info()->offset_first_element_in_bytes());
     const size_t num_classes = tensor.info()->dimension(0);
 
     classes_prob.resize(num_classes);
@@ -170,10 +168,28 @@ bool TopNPredictionsAccessor::access_tensor(ITensor &tensor)
     for(size_t i = 0; i < _top_n; ++i)
     {
         _output_stream << std::fixed << std::setprecision(4)
-                       << classes_prob[index.at(i)]
+                       << +classes_prob[index.at(i)]
                        << " - [id = " << index.at(i) << "]"
                        << ", " << _labels[index.at(i)] << std::endl;
     }
+}
+
+bool TopNPredictionsAccessor::access_tensor(ITensor &tensor)
+{
+    ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(&tensor, 1, DataType::F32, DataType::QASYMM8);
+    ARM_COMPUTE_ERROR_ON(_labels.size() != tensor.info()->dimension(0));
+
+    switch(tensor.info()->data_type())
+    {
+        case DataType::QASYMM8:
+            access_predictions_tensor<uint8_t>(tensor);
+            break;
+        case DataType::F32:
+            access_predictions_tensor<float>(tensor);
+            break;
+        default:
+            ARM_COMPUTE_ERROR("NOT SUPPORTED!");
+    }
 
     return false;
 }
diff --git a/utils/GraphUtils.h b/utils/GraphUtils.h
index ae6ecad881..e97bbf1c49 100644
--- a/utils/GraphUtils.h
+++ b/utils/GraphUtils.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017 ARM Limited.
+ * Copyright (c) 2017-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -128,6 +128,9 @@ public:
     bool access_tensor(ITensor &tensor) override;
 
 private:
+    template <typename T>
+    void access_predictions_tensor(ITensor &tensor);
+
     std::vector<std::string> _labels;
     std::ostream            &_output_stream;
     size_t                   _top_n;
diff --git a/utils/Utils.h b/utils/Utils.h
index 01e5137669..ff4c4c99fd 100644
--- a/utils/Utils.h
+++ b/utils/Utils.h
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2016, 2018 ARM Limited.
+ * Copyright (c) 2016-2018 ARM Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -138,6 +138,7 @@ inline std::string get_typestring(DataType data_type)
     switch(data_type)
     {
         case DataType::U8:
+        case DataType::QASYMM8:
             return no_endianness + "u" + support::cpp11::to_string(sizeof(uint8_t));
         case DataType::S8:
             return no_endianness + "i" + support::cpp11::to_string(sizeof(int8_t));