COMPMID-674 - Create Google InceptionV3 example

Change-Id: I389e0d4104b7dde60b7cdd612a83f3328517e44c
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/115804
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>

2 files changed, 21 insertions, 66 deletions

diff --git a/src/graph/nodes/BranchLayer.cpp b/src/graph/nodes/BranchLayer.cpp
index 6352bfc1e3..7a20a565b8 100644
--- a/src/graph/nodes/BranchLayer.cpp
+++ b/src/graph/nodes/BranchLayer.cpp
@@ -37,46 +37,6 @@
 
 using namespace arm_compute::graph;
 
-namespace
-{
-void depth_concatenate_output_info(ITensorInfo *info, ITensorInfo *sub_tensor_info)
-{
-    ARM_COMPUTE_ERROR_ON(info == nullptr);
-    ARM_COMPUTE_ERROR_ON(sub_tensor_info == nullptr);
-
-    TensorShape        info_shape            = info->tensor_shape();
-    const TensorShape &sub_tensor_info_shape = sub_tensor_info->tensor_shape();
-
-    // Update parent info and valid region
-    if(info_shape.total_size() == 0)
-    {
-        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->quantization_info());
-        info->set_valid_region(sub_tensor_info->valid_region());
-    }
-    else
-    {
-        ARM_COMPUTE_ERROR_ON(info->num_channels() != sub_tensor_info->num_channels());
-        ARM_COMPUTE_ERROR_ON(info->data_type() != sub_tensor_info->data_type());
-        ARM_COMPUTE_ERROR_ON(info->fixed_point_position() != sub_tensor_info->fixed_point_position());
-
-        // Concatenate depth
-        ARM_COMPUTE_ERROR_ON(info_shape.x() != sub_tensor_info_shape.x());
-        ARM_COMPUTE_ERROR_ON(info_shape.y() != sub_tensor_info_shape.y());
-        info_shape.set(2, info_shape.z() + sub_tensor_info_shape.z());
-        info->set_tensor_shape(info_shape);
-
-        // Update valid region
-        arm_compute::ValidRegion info_valid_region = info->valid_region();
-        info_valid_region.shape.set(2, info_shape.z());
-        arm_compute::ValidRegion updated_region = arm_compute::intersect_valid_regions(info_valid_region, sub_tensor_info->valid_region());
-        info->set_valid_region(updated_region);
-    }
-}
-} // namespace
-
 /** Branch function */
 class BranchFunction final : public arm_compute::IFunction
 {
@@ -117,9 +77,8 @@ std::unique_ptr<arm_compute::IFunction> BranchLayer::instantiate_node(GraphConte
     // Create branch function
     auto func = arm_compute::support::cpp14::make_unique<BranchFunction>();
 
-    // Track output SubTensorInfo and depth
-    TensorInfo out_info;
-    int        depth = 0;
+    // Track output depth
+    int depth = 0;
 
     // Constuct all sub-graphs given the input/output
     for(auto &sg : _sub_graphs)
@@ -143,10 +102,13 @@ std::unique_ptr<arm_compute::IFunction> BranchLayer::instantiate_node(GraphConte
         // Create output sub-tensor
         if(!sg->has_output())
         {
-            ARM_COMPUTE_ERROR_ON(dynamic_cast<Tensor *>(output) == nullptr);
-            out = arm_compute::support::cpp14::make_unique<SubTensor>(*dynamic_cast<Tensor *>(output),
-                                                                      output->tensor()->info()->tensor_shape(),
-                                                                      Coordinates(0, 0, depth));
+            ARM_COMPUTE_ERROR_ON((dynamic_cast<Tensor *>(output) == nullptr) && (dynamic_cast<SubTensor *>(output) == nullptr));
+
+            out = arm_compute::support::cpp14::make_unique<SubTensor>(output->tensor(),
+                                                                      TensorShape(),
+                                                                      Coordinates(0, 0, depth),
+                                                                      output->target(),
+                                                                      true);
             out_sub_tensor = dynamic_cast<SubTensor *>(out.get());
         }
 
@@ -161,17 +123,8 @@ std::unique_ptr<arm_compute::IFunction> BranchLayer::instantiate_node(GraphConte
         {
             ARM_COMPUTE_ERROR_ON(out_sub_tensor->tensor() == nullptr);
             depth += out_sub_tensor->tensor()->info()->tensor_shape()[2];
-            depth_concatenate_output_info(&out_info, out_sub_tensor->tensor()->info());
         }
     }
 
-    // Auto-init output
-    arm_compute::auto_init_if_empty(*output->tensor()->info(),
-                                    out_info.tensor_shape(),
-                                    out_info.num_channels(),
-                                    out_info.data_type(),
-                                    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 53d06ea75f..f292b893ed 100644
--- a/src/graph/nodes/ConvolutionLayer.cpp
+++ b/src/graph/nodes/ConvolutionLayer.cpp
@@ -106,13 +106,16 @@ std::unique_ptr<arm_compute::IFunction> instantiate<TargetHint::OPENCL>(arm_comp
                                                                         const WeightsInfo    &weights_info,
                                                                         ConvolutionMethodHint conv_method)
 {
-    if(conv_method == ConvolutionMethodHint::GEMM)
+    if((conv_method == ConvolutionMethodHint::DIRECT)
+       && arm_compute::CLDirectConvolutionLayer::validate(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info)) // NOLINT
     {
-        return instantiate_function<arm_compute::CLConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info, weights_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLDirectConvolutionLayer");
+        return instantiate_direct_function<arm_compute::CLDirectConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info);
     }
     else
     {
-        return instantiate_direct_function<arm_compute::CLDirectConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLConvolutionLayer");
+        return instantiate_function<arm_compute::CLConvolutionLayer, arm_compute::ICLTensor, TargetHint::OPENCL>(input, weights, biases, output, conv_info, weights_info);
     }
 }
 
@@ -122,13 +125,16 @@ std::unique_ptr<arm_compute::IFunction> instantiate<TargetHint::NEON>(arm_comput
                                                                       const WeightsInfo    &weights_info,
                                                                       ConvolutionMethodHint conv_method)
 {
-    if(conv_method == ConvolutionMethodHint::GEMM)
+    if((conv_method == ConvolutionMethodHint::DIRECT)
+       && arm_compute::NEDirectConvolutionLayer::validate(input->info(), weights->info(), biases != nullptr ? biases->info() : nullptr, output->info(), conv_info)) // NOLINT
     {
-        return instantiate_function<arm_compute::NEConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info, weights_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEDirectConvolutionLayer");
+        return instantiate_direct_function<arm_compute::NEDirectConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info);
     }
     else
     {
-        return instantiate_direct_function<arm_compute::NEDirectConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info);
+        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEConvolutionLayer");
+        return instantiate_function<arm_compute::NEConvolutionLayer, arm_compute::ITensor, TargetHint::NEON>(input, weights, biases, output, conv_info, weights_info);
     }
 }
 } // namespace
@@ -258,12 +264,10 @@ std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_convolutio
     std::unique_ptr<arm_compute::IFunction> func;
     if(_target_hint == TargetHint::OPENCL)
     {
-        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLConvolutionLayer");
         func = instantiate<TargetHint::OPENCL>(input, _weights.tensor(), _biases.tensor(), output, _conv_info, _weights_info, conv_method_hint);
     }
     else
     {
-        ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEConvolutionLayer");
         func = instantiate<TargetHint::NEON>(input, _weights.tensor(), _biases.tensor(), output, _conv_info, _weights_info, conv_method_hint);
     }
     return func;
@@ -325,12 +329,10 @@ std::unique_ptr<arm_compute::IFunction> ConvolutionLayer::instantiate_grouped_co
         // Instantiate convolution function
         if(_target_hint == TargetHint::OPENCL)
         {
-            ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating CLConvolutionLayer");
             func = instantiate<TargetHint::OPENCL>(_is[i].tensor(), _ws[i].tensor(), _bs[i].tensor(), _os[i].tensor(), _conv_info, _weights_info, conv_method_hint);
         }
         else
         {
-            ARM_COMPUTE_LOG_GRAPH_INFO("Instantiating NEConvolutionLayer");
             func = instantiate<TargetHint::NEON>(_is[i].tensor(), _ws[i].tensor(), _bs[i].tensor(), _os[i].tensor(), _conv_info, _weights_info, conv_method_hint);
         }