dnn: prevent unloading of InferenceEngine plugins

Add lightweight IE hardware targets checks

nGraph: Concat with paddings

Enable more nGraph tests

Restore FP32->FP16 for GPU plugin of IE

try to fix buildbot

Use lightweight IE targets check only starts from R4

LICENSE

@@ -7,7 +7,7 @@ copy or use the software.
                For Open Source Computer Vision Library
                        (3-clause BSD License)
 
-Copyright (C) 2000-2019, Intel Corporation, all rights reserved.
+Copyright (C) 2000-2020, Intel Corporation, all rights reserved.
 Copyright (C) 2009-2011, Willow Garage Inc., all rights reserved.
 Copyright (C) 2009-2016, NVIDIA Corporation, all rights reserved.
 Copyright (C) 2010-2013, Advanced Micro Devices, Inc., all rights reserved.

cmake/OpenCVDetectInferenceEngine.cmake

@@ -94,9 +94,9 @@ endif()
 
 if(INF_ENGINE_TARGET)
   if(NOT INF_ENGINE_RELEASE)
-    message(WARNING "InferenceEngine version have not been set, 2019R3 will be used by default. Set INF_ENGINE_RELEASE variable if you experience build errors.")
+    message(WARNING "InferenceEngine version has not been set, 2020.1 will be used by default. Set INF_ENGINE_RELEASE variable if you experience build errors.")
   endif()
-  set(INF_ENGINE_RELEASE "2019030000" CACHE STRING "Force IE version, should be in form YYYYAABBCC (e.g. 2018R2.0.2 -> 2018020002)")
+  set(INF_ENGINE_RELEASE "2020010000" CACHE STRING "Force IE version, should be in form YYYYAABBCC (e.g. 2020.1.0.2 -> 2020010002)")
   set_target_properties(${INF_ENGINE_TARGET} PROPERTIES
     INTERFACE_COMPILE_DEFINITIONS "HAVE_INF_ENGINE=1;INF_ENGINE_RELEASE=${INF_ENGINE_RELEASE}"
   )

modules/core/include/opencv2/core/version.hpp

@@ -8,7 +8,7 @@
 #define CV_VERSION_MAJOR    4
 #define CV_VERSION_MINOR    2
 #define CV_VERSION_REVISION 0
-#define CV_VERSION_STATUS   ""
+#define CV_VERSION_STATUS   "-openvino"
 
 #define CVAUX_STR_EXP(__A)  #__A
 #define CVAUX_STR(__A)      CVAUX_STR_EXP(__A)

modules/dnn/src/dnn.cpp

@@ -109,6 +109,22 @@ public:
 #ifdef HAVE_INF_ENGINE
     static inline bool checkIETarget(Target target)
     {
+#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R3)
+        // Lightweight detection
+        const std::vector<std::string> devices = getCore().GetAvailableDevices();
+        for (std::vector<std::string>::const_iterator i = devices.begin(); i != devices.end(); ++i)
+        {
+            if (std::string::npos != i->find("MYRIAD") && target == DNN_TARGET_MYRIAD)
+                return true;
+            else if (std::string::npos != i->find("FPGA") && target == DNN_TARGET_FPGA)
+                return true;
+            else if (std::string::npos != i->find("CPU") && target == DNN_TARGET_CPU)
+                return true;
+            else if (std::string::npos != i->find("GPU") && (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
+                return true;
+        }
+        return false;
+#else
         cv::dnn::Net net;
         cv::dnn::LayerParams lp;
         lp.set("kernel_size", 1);
@@ -132,6 +148,7 @@ public:
             return false;
         }
         return true;
+#endif
     }
 #endif
 

modules/dnn/src/graph_simplifier.cpp

@@ -0,0 +1,207 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+// Copyright (C) 2020, Intel Corporation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#include "precomp.hpp"
+
+#include "graph_simplifier.hpp"
+
+#include <queue>
+
+namespace cv { namespace dnn {
+
+Subgraph::~Subgraph() {}
+
+int Subgraph::addNodeToMatch(const std::string& op, int input_0, int input_1,
+                             int input_2, int input_3)
+{
+    int nodeInputs[] = {input_0, input_1, input_2, input_3};
+    int numInputs = 0;
+    for (int i = 0; i < 4; ++i)
+    {
+        numInputs += (int)(nodeInputs[i] != -1);
+    }
+    return addNodeToMatch(op, std::vector<int>(&nodeInputs[0], &nodeInputs[0] + numInputs));
+}
+
+int Subgraph::addNodeToMatch(const std::string& op, const std::vector<int>& inputs_)
+{
+    for (int i = 0; i < inputs_.size(); ++i)
+    {
+        CV_Assert(inputs_[i] < (int)nodes.size());
+    }
+    nodes.push_back(op);
+    inputs.push_back(inputs_);
+    return nodes.size() - 1;
+}
+
+void Subgraph::setFusedNode(const std::string& op, int input_0, int input_1,
+                            int input_2, int input_3, int input_4, int input_5)
+{
+    int nodeInputs[] = {input_0, input_1, input_2, input_3, input_4, input_5};
+    int numInputs = 0;
+    for (int i = 0; i < 6; ++i)
+    {
+        CV_Assert(nodeInputs[i] < (int)nodes.size());
+        numInputs += (int)(nodeInputs[i] != -1);
+    }
+    setFusedNode(op, std::vector<int>(&nodeInputs[0], &nodeInputs[0] + numInputs));
+}
+
+void Subgraph::setFusedNode(const std::string& op, const std::vector<int>& inputs_)
+{
+    fusedNodeInputs = inputs_;
+    fusedNodeOp = op;
+}
+
+int Subgraph::getInputNodeId(const Ptr<ImportGraphWrapper>& net,
+                             const Ptr<ImportNodeWrapper>& node,
+                             int inpId)
+{
+    CV_Assert(inpId < node->getNumInputs());
+    std::string name = node->getInputName(inpId);
+    // If operation produces several tensors, they are specified by index
+    // after ':' character. In example, "input:0".
+    name = name.substr(0, name.rfind(':'));
+    const int numNodes = net->getNumNodes();
+    for (int i = 0; i < numNodes; ++i)
+    {
+        if (net->getNodeName(i) == name)
+            return i;
+    }
+    CV_Error(Error::StsParseError, "Input node with name " + name + " not found");
+}
+
+bool Subgraph::match(const Ptr<ImportGraphWrapper>& net, int nodeId,
+                     std::vector<int>& matchedNodesIds,
+                     std::vector<int>& targetNodesIds)
+{
+    matchedNodesIds.clear();
+    targetNodesIds.clear();
+
+    std::queue<int> nodesToMatch;
+    std::queue<int> targetNodes;
+    nodesToMatch.push(nodeId);
+    targetNodes.push(nodes.size() - 1);
+    while (!nodesToMatch.empty())
+    {
+        int nodeToMatch = nodesToMatch.front();
+        int targetNodeId = targetNodes.front();
+        nodesToMatch.pop();
+        targetNodes.pop();
+
+        if (std::find(matchedNodesIds.begin(), matchedNodesIds.end(), nodeToMatch) !=
+            matchedNodesIds.end())
+            continue;
+
+        const Ptr<ImportNodeWrapper> node = net->getNode(nodeToMatch);
+        if (node->getType() != nodes[targetNodeId])
+            return false;
+
+        std::vector<int>& inputNodes = inputs[targetNodeId];
+        if (inputNodes.size() != node->getNumInputs())
+            return false;
+
+        for (int j = 0; j < inputNodes.size(); ++j)
+        {
+            if (nodes[inputNodes[j]].empty())  // Unknown input node type.
+                continue;
+            nodeId = getInputNodeId(net, node, j);
+            const Ptr<ImportNodeWrapper> inpNode = net->getNode(nodeId);
+            if (inpNode->getType() != "Const")
+            {
+                nodesToMatch.push(nodeId);
+                targetNodes.push(inputNodes[j]);
+            }
+            else if (nodes[inputNodes[j]] != "Const")
+                return false;
+        }
+        matchedNodesIds.push_back(nodeToMatch);
+        targetNodesIds.push_back(targetNodeId);
+    }
+
+    const int n = matchedNodesIds.size();
+    std::vector<std::pair<int, int> > elements(n);
+    for (int i = 0; i < n; ++i)
+        elements[i] = std::make_pair(matchedNodesIds[i], targetNodesIds[i]);
+    std::sort(elements.begin(), elements.end());
+    for (int i = 0; i < n; ++i)
+    {
+        matchedNodesIds[i] = elements[i].first;
+        targetNodesIds[i] = elements[i].second;
+    }
+    return true;
+}
+
+void Subgraph::replace(const Ptr<ImportGraphWrapper>& net, const std::vector<int>& matchedNodesIds,
+                       const std::vector<int>& targetNodesIds)
+{
+    // Extract names of input nodes.
+    std::vector<std::string> inputsNames(fusedNodeInputs.size());
+    for (int i = 0; i < fusedNodeInputs.size(); ++i)
+    {
+        std::string inpName;
+        // Find input node name looking at inputs of fused nodes.
+        for (int j = 0; j < matchedNodesIds.size() && inpName.empty(); ++j)
+        {
+            Ptr<ImportNodeWrapper> node = net->getNode(matchedNodesIds[j]);
+            std::vector<int>& inpIndices = inputs[targetNodesIds[j]];
+
+            CV_Assert(node->getNumInputs() == inpIndices.size());
+            for (int k = 0; k < inpIndices.size(); ++k)
+            {
+                if (inpIndices[k] == fusedNodeInputs[i])
+                {
+                    inpName = node->getInputName(k);
+                    break;
+                }
+            }
+        }
+        CV_Assert(!inpName.empty());
+        inputsNames[i] = inpName;
+    }
+
+    // Remove matched nodes except the last one. Indices in ascending order are expected.
+    Ptr<ImportNodeWrapper> node = net->getNode(matchedNodesIds.back());
+    for (int i = matchedNodesIds.size() - 2; i >= 0; --i)
+        net->removeNode(matchedNodesIds[i]);
+
+    // Modify the last node to be a fused one.
+    node->setType(fusedNodeOp);
+    node->setInputNames(inputsNames);
+
+    std::vector<Ptr<ImportNodeWrapper> > inputNodes(inputsNames.size());
+    for (int i = 0; i < inputsNames.size(); ++i)
+    {
+        inputNodes[i] = net->getNode(getInputNodeId(net, node, i));
+    }
+    finalize(net, node, inputNodes);
+}
+
+void Subgraph::finalize(const Ptr<ImportGraphWrapper>& net,
+                        const Ptr<ImportNodeWrapper>& fusedNode,
+                        std::vector<Ptr<ImportNodeWrapper> >& inputs) {}
+
+void simplifySubgraphs(const Ptr<ImportGraphWrapper>& net,
+                       const std::vector<Ptr<Subgraph> >& patterns)
+{
+    int numNodes = net->getNumNodes();
+    std::vector<int> matchedNodesIds, targetNodesIds;
+    for (int i = 0; i < numNodes; ++i)
+    {
+        for (int j = 0; j < patterns.size(); ++j)
+        {
+            if (patterns[j]->match(net, i, matchedNodesIds, targetNodesIds))
+            {
+                patterns[j]->replace(net, matchedNodesIds, targetNodesIds);
+                numNodes -= matchedNodesIds.size() - 1;  // #matchedNodes removed and one added.
+                break;
+            }
+        }
+    }
+}
+
+}}  // namespace cv::dnn

modules/dnn/src/graph_simplifier.hpp

@@ -0,0 +1,100 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+// Copyright (C) 2020, Intel Corporation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#ifndef __OPENCV_DNN_GRAPH_SIMPLIFIER_HPP__
+#define __OPENCV_DNN_GRAPH_SIMPLIFIER_HPP__
+
+#include <string>
+
+#include <opencv2/core.hpp>
+
+namespace cv { namespace dnn {
+
+class ImportNodeWrapper
+{
+public:
+    virtual ~ImportNodeWrapper() {};
+
+    virtual int getNumInputs() const = 0;
+
+    virtual std::string getInputName(int idx) const = 0;
+
+    virtual std::string getType() const = 0;
+
+    virtual void setType(const std::string& type) = 0;
+
+    virtual void setInputNames(const std::vector<std::string>& inputs) = 0;
+};
+
+class ImportGraphWrapper
+{
+public:
+    virtual ~ImportGraphWrapper() {};
+
+    virtual Ptr<ImportNodeWrapper> getNode(int idx) const = 0;
+
+    virtual int getNumNodes() const = 0;
+
+    virtual std::string getNodeName(int idx) const = 0;
+
+    virtual void removeNode(int idx) = 0;
+};
+
+class Subgraph  // Interface to match and replace subgraphs.
+{
+public:
+    virtual ~Subgraph();
+
+    // Add a node to be matched in the origin graph. Specify ids of nodes that
+    // are expected to be inputs. Returns id of a newly added node.
+    // TODO: Replace inputs to std::vector<int> in C++11
+    int addNodeToMatch(const std::string& op, int input_0 = -1, int input_1 = -1,
+                       int input_2 = -1, int input_3 = -1);
+
+    int addNodeToMatch(const std::string& op, const std::vector<int>& inputs_);
+
+    // Specify resulting node. All the matched nodes in subgraph excluding
+    // input nodes will be fused into this single node.
+    // TODO: Replace inputs to std::vector<int> in C++11
+    void setFusedNode(const std::string& op, int input_0 = -1, int input_1 = -1,
+                      int input_2 = -1, int input_3 = -1, int input_4 = -1,
+                      int input_5 = -1);
+
+    void setFusedNode(const std::string& op, const std::vector<int>& inputs_);
+
+    static int getInputNodeId(const Ptr<ImportGraphWrapper>& net,
+                              const Ptr<ImportNodeWrapper>& node,
+                              int inpId);
+
+    // Match TensorFlow subgraph starting from <nodeId> with a set of nodes to be fused.
+    // Const nodes are skipped during matching. Returns true if nodes are matched and can be fused.
+    virtual bool match(const Ptr<ImportGraphWrapper>& net, int nodeId,
+                       std::vector<int>& matchedNodesIds,
+                       std::vector<int>& targetNodesIds);
+
+    // Fuse matched subgraph.
+    void replace(const Ptr<ImportGraphWrapper>& net, const std::vector<int>& matchedNodesIds,
+                 const std::vector<int>& targetNodesIds);
+
+    virtual void finalize(const Ptr<ImportGraphWrapper>& net,
+                          const Ptr<ImportNodeWrapper>& fusedNode,
+                          std::vector<Ptr<ImportNodeWrapper> >& inputs);
+
+private:
+    std::vector<std::string> nodes;         // Nodes to be matched in the origin graph.
+    std::vector<std::vector<int> > inputs;  // Connections of an every node to it's inputs.
+
+    std::string fusedNodeOp;           // Operation name of resulting fused node.
+    std::vector<int> fusedNodeInputs;  // Inputs of fused node.
+};
+
+void simplifySubgraphs(const Ptr<ImportGraphWrapper>& net,
+                       const std::vector<Ptr<Subgraph> >& patterns);
+
+}}  // namespace dnn, namespace cv
+
+#endif  // __OPENCV_DNN_GRAPH_SIMPLIFIER_HPP__

modules/dnn/src/ie_ngraph.cpp

@@ -168,21 +168,26 @@ void InfEngineNgraphNet::init(Target targetId)
 {
     if (!hasNetOwner)
     {
-        if (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) {
+        if (targetId == DNN_TARGET_OPENCL_FP16)
+        {
             auto nodes = ngraph_function->get_ordered_ops();
-            for (auto& node : nodes) {
+            for (auto& node : nodes)
+            {
                 auto parameter = std::dynamic_pointer_cast<ngraph::op::Parameter>(node);
-                if (parameter && parameter->get_element_type() == ngraph::element::f32) {
+                if (parameter && parameter->get_element_type() == ngraph::element::f32)
+                {
                     parameter->set_element_type(ngraph::element::f16);
                 }
                 auto constant = std::dynamic_pointer_cast<ngraph::op::Constant>(node);
-                if (constant && constant->get_element_type() == ngraph::element::f32) {
-                    auto data = constant->get_vector<float>();
-                    std::vector<ngraph::float16> new_data(data.size());
-                    for (size_t i = 0; i < data.size(); ++i) {
-                        new_data[i] = ngraph::float16(data[i]);
-                    }
-                    auto new_const = std::make_shared<ngraph::op::Constant>(ngraph::element::f16, constant->get_shape(), new_data);
+                if (constant && constant->get_element_type() == ngraph::element::f32)
+                {
+                    const float* floatsData = constant->get_data_ptr<float>();
+                    size_t total = ngraph::shape_size(constant->get_shape());
+                    Mat floats(1, total, CV_32F, (void*)floatsData);
+                    Mat halfs;
+                    cv::convertFp16(floats, halfs);
+
+                    auto new_const = std::make_shared<ngraph::op::Constant>(ngraph::element::f16, constant->get_shape(), halfs.data);
                     new_const->set_friendly_name(constant->get_friendly_name());
                     ngraph::replace_node(constant, new_const);
                 }
@@ -318,7 +323,14 @@ void InfEngineNgraphNet::initPlugin(InferenceEngine::CNNNetwork& net)
             }
             // Some of networks can work without a library of extra layers.
             // OpenCV fallbacks as extensions.
-            ie.AddExtension(std::make_shared<InfEngineExtension>(), "CPU");
+            try
+            {
+                ie.AddExtension(std::make_shared<InfEngineExtension>(), "CPU");
+            }
+            catch(const std::exception& e)
+            {
+                CV_LOG_INFO(NULL, "DNN-IE: Can't register OpenCV custom layers extension: " << e.what());
+            }
 #ifndef _WIN32
             // Limit the number of CPU threads.
             if (device_name == "CPU")

modules/dnn/src/layers/concat_layer.cpp

@@ -114,7 +114,8 @@ public:
         return backendId == DNN_BACKEND_OPENCV ||
                backendId == DNN_BACKEND_CUDA ||
                (backendId == DNN_BACKEND_HALIDE && haveHalide() && axis == 1 && !padding) ||  // By channels
-               ((backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && haveInfEngine() && !padding) ||
+               (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && haveInfEngine() && !padding) ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH ||
                (backendId == DNN_BACKEND_VKCOM && haveVulkan() && !padding);
     }
 
@@ -351,14 +352,45 @@ public:
     virtual Ptr<BackendNode> initNgraph(const std::vector<Ptr<BackendWrapper> >& inputs,
                                         const std::vector<Ptr<BackendNode> >& nodes) CV_OVERRIDE
     {
+        InferenceEngine::DataPtr data = ngraphDataNode(inputs[0]);
+        const int numDims = data->getDims().size();
+        const int cAxis = clamp(axis, numDims);
+        std::vector<size_t> maxDims(numDims, 0);
+
         CV_Assert(inputs.size() == nodes.size());
         ngraph::NodeVector inp_nodes;
-        for (auto& node : nodes) {
-            inp_nodes.push_back(node.dynamicCast<InfEngineNgraphNode>()->node);
-        }
+        for (int i = 0; i < nodes.size(); ++i)
+        {
+            inp_nodes.push_back(nodes[i].dynamicCast<InfEngineNgraphNode>()->node);
 
-        InferenceEngine::DataPtr data = ngraphDataNode(inputs[0]);
-        auto concat = std::make_shared<ngraph::op::Concat>(inp_nodes, clamp(axis, data->getDims().size()));
+            std::vector<size_t> inpShape = ngraphDataNode(inputs[i])->getDims();
+            for (int i = 0; i < numDims; ++i)
+                maxDims[i] = std::max(maxDims[i], inpShape[i]);
+        }
+        for (int i = 0; i < inp_nodes.size(); ++i)
+        {
+            bool needPadding = false;
+            std::vector<size_t> inpShape = ngraphDataNode(inputs[i])->getDims();
+            std::vector<int64_t> begins(inpShape.size(), 0), ends(inpShape.size(), 0);
+            for (int j = 0; j < inpShape.size(); ++j)
+            {
+                if (j != cAxis && inpShape[j] != maxDims[j])
+                {
+                    needPadding = true;
+                    begins[j] = static_cast<int64_t>((maxDims[j] - inpShape[j]) / 2);
+                    ends[j] = static_cast<int64_t>(maxDims[j] - inpShape[j] - begins[j]);
+                }
+            }
+            if (needPadding)
+            {
+                inp_nodes[i] = std::make_shared<ngraph::op::v1::Pad>(
+                    inp_nodes[i],
+                    std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{begins.size()}, begins.data()),
+                    std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{ends.size()}, ends.data()),
+                    ngraph::op::PadMode::CONSTANT);
+            }
+        }
+        auto concat = std::make_shared<ngraph::op::Concat>(inp_nodes, cAxis);
         return Ptr<BackendNode>(new InfEngineNgraphNode(concat));
     }
 #endif  // HAVE_DNN_NGRAPH

modules/dnn/src/layers/lrn_layer.cpp

@@ -103,7 +103,7 @@ public:
             return bias == (int)bias;
         }
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) {
-            return type == CHANNEL_NRM && bias == (int)bias;
+            return bias == (int)bias;
         }
         return backendId == DNN_BACKEND_OPENCV ||
                backendId == DNN_BACKEND_CUDA ||
@@ -471,7 +471,15 @@ public:
             alphaSize *= (type == SPATIAL_NRM ? size*size : size);
 
         auto& ieInpNode = nodes[0].dynamicCast<InfEngineNgraphNode>()->node;
-        auto lrn = std::make_shared<ngraph::op::LRN>(ieInpNode, (double)alphaSize, (double)beta, (double)bias, (size_t)size);
+        std::vector<int64_t> axes;
+        if (type != SPATIAL_NRM) {
+            axes = {1};
+        } else {
+            axes.resize(ieInpNode->get_shape().size() - 2);
+            std::iota(axes.begin(), axes.end(), 2);
+        }
+        auto ngraph_axes = std::make_shared<ngraph::op::Constant>(ngraph::element::i64, ngraph::Shape{axes.size()}, axes.data());
+        auto lrn = std::make_shared<ngraph::op::LRN>(ieInpNode, ngraph_axes, alphaSize, beta, bias, size);
         return Ptr<BackendNode>(new InfEngineNgraphNode(lrn));
     }
 #endif  // HAVE_DNN_NGRAPH

modules/dnn/src/layers/mvn_layer.cpp

@@ -119,8 +119,10 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
 #ifdef HAVE_INF_ENGINE
-        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
+        if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019)
             return !zeroDev && (preferableTarget != DNN_TARGET_MYRIAD || eps <= 1e-7f);
+        else if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
+            return true;
         else
 #endif  // HAVE_INF_ENGINE
             return backendId == DNN_BACKEND_OPENCV;

modules/dnn/src/layers/normalize_bbox_layer.cpp

@@ -75,7 +75,10 @@ public:
             if (pnorm != 2)
                 return false;
 
-            return preferableTarget == DNN_TARGET_MYRIAD ? !acrossSpatial : startAxis == 1;
+            if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && preferableTarget == DNN_TARGET_MYRIAD)
+                return !acrossSpatial;
+
+            return startAxis == 1;
         }
         return backendId == DNN_BACKEND_OPENCV ||
                (backendId == DNN_BACKEND_CUDA && (pnorm == 1 || pnorm == 2));
@@ -373,7 +376,6 @@ public:
         }
         else
         {
-            // weight->get_shape().size() > 1 ~> channel_shared = false
             weight = std::make_shared<ngraph::op::Constant>(
                                       ngraph::element::f32, ngraph::Shape(shape), blobs[0].data);
         }

modules/dnn/src/layers/pooling_layer.cpp

@@ -203,7 +203,7 @@ public:
 #endif
         }
         else if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) {
-            return type != STOCHASTIC;
+            return !computeMaxIdx && type != STOCHASTIC;
         }
         else
         {

modules/dnn/src/onnx/onnx_graph_simplifier.cpp

@@ -0,0 +1,157 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+// Copyright (C) 2020, Intel Corporation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#include "../precomp.hpp"
+
+#include "../graph_simplifier.hpp"
+#include "onnx_graph_simplifier.hpp"
+
+#include <queue>
+
+namespace cv { namespace dnn {
+CV__DNN_INLINE_NS_BEGIN
+
+// This wrapper can behave differently for fake input nodes and real graph nodes.
+class ONNXNodeWrapper : public ImportNodeWrapper
+{
+public:
+    ONNXNodeWrapper(opencv_onnx::NodeProto* _node = 0) : node(_node) {}
+
+    virtual int getNumInputs() const CV_OVERRIDE
+    {
+        return node ? node->input_size() : 0;
+    }
+
+    virtual std::string getInputName(int idx) const CV_OVERRIDE
+    {
+        CV_Assert_N(node, idx < node->input_size());
+        return node->input(idx);
+    }
+
+    virtual std::string getType() const CV_OVERRIDE
+    {
+        return node ? node->op_type() : "";
+    }
+
+    virtual void setType(const std::string& type) CV_OVERRIDE
+    {
+        CV_Assert(node);
+        node->set_op_type(type);
+    }
+
+    virtual void setInputNames(const std::vector<std::string>& inputs) CV_OVERRIDE
+    {
+        CV_Assert(node);
+        node->clear_input();
+        for (int i = 0; i < inputs.size(); ++i)
+            node->add_input(inputs[i]);
+    }
+
+    opencv_onnx::NodeProto* node;
+};
+
+// ONNX graph's inputs are separate from nodes so we index them before the rest of nodes.
+class ONNXGraphWrapper : public ImportGraphWrapper
+{
+public:
+    ONNXGraphWrapper(opencv_onnx::GraphProto& _net) : net(_net)
+    {
+        numInputs = net.input_size();
+    }
+
+    virtual Ptr<ImportNodeWrapper> getNode(int idx) const CV_OVERRIDE
+    {
+        opencv_onnx::NodeProto* node = 0;
+        if (idx >= numInputs)
+            node = net.mutable_node(idx - numInputs);
+        return makePtr<ONNXNodeWrapper>(node);
+    }
+
+    virtual int getNumNodes() const CV_OVERRIDE
+    {
+        return numInputs + net.node_size();
+    }
+
+    virtual std::string getNodeName(int idx) const CV_OVERRIDE
+    {
+        if (idx < numInputs)
+            return net.input(idx).name();
+        else
+            return net.node(idx - numInputs).output(0);
+    }
+
+    virtual void removeNode(int idx) CV_OVERRIDE
+    {
+        CV_Assert(idx >= numInputs);
+        net.mutable_node()->DeleteSubrange(idx - numInputs, 1);
+    }
+
+private:
+    int numInputs;
+    opencv_onnx::GraphProto& net;
+};
+
+class SoftMaxSubgraph : public Subgraph
+{
+public:
+    SoftMaxSubgraph() : axis(1)
+    {
+        int input = addNodeToMatch("");
+        int inpExp = addNodeToMatch("Exp", input);
+        int sum = addNodeToMatch("ReduceSum", inpExp);
+        addNodeToMatch("Div", inpExp, sum);
+        setFusedNode("Softmax", input);
+    }
+
+    virtual bool match(const Ptr<ImportGraphWrapper>& net, int nodeId,
+                       std::vector<int>& matchedNodesIds,
+                       std::vector<int>& targetNodesIds) CV_OVERRIDE
+    {
+        if (Subgraph::match(net, nodeId, matchedNodesIds, targetNodesIds))
+        {
+            Ptr<ImportNodeWrapper> sum = net->getNode(matchedNodesIds[1]);
+            opencv_onnx::NodeProto* node = sum.dynamicCast<ONNXNodeWrapper>()->node;
+
+            for (int i = 0; i < node->attribute_size(); i++)
+            {
+                opencv_onnx::AttributeProto attr = node->attribute(i);
+                if (attr.name() != "axes")
+                    continue;
+                if (attr.ints_size() != 1)
+                    CV_Error(Error::StsNotImplemented, format("Unexpected number of axes: %d", attr.ints_size()));
+                axis = attr.ints(0);
+                return true;
+            }
+            CV_Error(Error::StsNotImplemented, "Missed axes attribute");
+        }
+        return false;
+    }
+
+    virtual void finalize(const Ptr<ImportGraphWrapper>&,
+                          const Ptr<ImportNodeWrapper>& fusedNode,
+                          std::vector<Ptr<ImportNodeWrapper> >&) CV_OVERRIDE
+    {
+        opencv_onnx::NodeProto* node = fusedNode.dynamicCast<ONNXNodeWrapper>()->node;
+        opencv_onnx::AttributeProto* attr = node->add_attribute();
+        attr->set_name("axis");
+        attr->set_i(axis);
+    }
+
+private:
+    int axis;
+};
+
+void simplifySubgraphs(opencv_onnx::GraphProto& net)
+{
+    std::vector<Ptr<Subgraph> > subgraphs;
+    subgraphs.push_back(makePtr<SoftMaxSubgraph>());
+
+    simplifySubgraphs(Ptr<ImportGraphWrapper>(new ONNXGraphWrapper(net)), subgraphs);
+}
+
+CV__DNN_INLINE_NS_END
+}}  // namespace cv::dnn

modules/dnn/src/onnx/onnx_graph_simplifier.hpp

@@ -0,0 +1,30 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at http://opencv.org/license.html.
+
+// Copyright (C) 2020, Intel Corporation, all rights reserved.
+// Third party copyrights are property of their respective owners.
+
+#ifndef __OPENCV_DNN_ONNX_SIMPLIFIER_HPP__
+#define __OPENCV_DNN_ONNX_SIMPLIFIER_HPP__
+
+#include "../precomp.hpp"
+
+#if defined(__GNUC__) && __GNUC__ >= 5
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wsuggest-override"
+#endif
+#include "opencv-onnx.pb.h"
+#if defined(__GNUC__) && __GNUC__ >= 5
+#pragma GCC diagnostic pop
+#endif
+
+namespace cv { namespace dnn {
+CV__DNN_INLINE_NS_BEGIN
+
+void simplifySubgraphs(opencv_onnx::GraphProto& net);
+
+CV__DNN_INLINE_NS_END
+}}  // namespace dnn, namespace cv
+
+#endif  // __OPENCV_DNN_ONNX_SIMPLIFIER_HPP__

modules/dnn/src/onnx/onnx_importer.cpp

@@ -26,6 +26,8 @@
 #pragma GCC diagnostic pop
 #endif
 
+#include "onnx_graph_simplifier.hpp"
+
 namespace cv {
 namespace dnn {
 CV__DNN_INLINE_NS_BEGIN
@@ -326,6 +328,9 @@ void ONNXImporter::populateNet(Net dstNet)
 {
     CV_Assert(model_proto.has_graph());
     opencv_onnx::GraphProto graph_proto = model_proto.graph();
+
+    simplifySubgraphs(graph_proto);
+
     std::map<std::string, Mat> constBlobs = getGraphTensors(graph_proto);
     // List of internal blobs shapes.
     std::map<std::string, MatShape> outShapes;

modules/dnn/src/op_inf_engine.cpp

@@ -564,16 +564,65 @@ static std::map<std::string, InferenceEngine::InferenceEnginePluginPtr>& getShar
     return sharedPlugins;
 }
 #else
-InferenceEngine::Core& getCore()
+static bool init_IE_plugins()
+{
+    // load and hold IE plugins
+    static InferenceEngine::Core* init_core = new InferenceEngine::Core();  // 'delete' is never called
+    (void)init_core->GetAvailableDevices();
+    return true;
+}
+static InferenceEngine::Core& create_IE_Core_instance()
 {
     static InferenceEngine::Core core;
     return core;
 }
+static InferenceEngine::Core& create_IE_Core_pointer()
+{
+    // load and hold IE plugins
+    static InferenceEngine::Core* core = new InferenceEngine::Core();  // 'delete' is never called
+    return *core;
+}
+InferenceEngine::Core& getCore()
+{
+    // to make happy memory leak tools use:
+    // - OPENCV_DNN_INFERENCE_ENGINE_HOLD_PLUGINS=0
+    // - OPENCV_DNN_INFERENCE_ENGINE_CORE_LIFETIME_WORKAROUND=0
+    static bool param_DNN_INFERENCE_ENGINE_HOLD_PLUGINS = utils::getConfigurationParameterBool("OPENCV_DNN_INFERENCE_ENGINE_HOLD_PLUGINS", true);
+    static bool init_IE_plugins_ = param_DNN_INFERENCE_ENGINE_HOLD_PLUGINS && init_IE_plugins(); CV_UNUSED(init_IE_plugins_);
+
+    static bool param_DNN_INFERENCE_ENGINE_CORE_LIFETIME_WORKAROUND =
+            utils::getConfigurationParameterBool("OPENCV_DNN_INFERENCE_ENGINE_CORE_LIFETIME_WORKAROUND",
+#ifdef _WIN32
+                true
+#else
+                false
+#endif
+            );
+    static InferenceEngine::Core& core = param_DNN_INFERENCE_ENGINE_CORE_LIFETIME_WORKAROUND
+            ? create_IE_Core_pointer()
+            : create_IE_Core_instance();
+    return core;
+}
 #endif
 
 #if !defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
 static bool detectMyriadX_()
 {
+#if INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R3)
+    // Lightweight detection
+    InferenceEngine::Core& ie = getCore();
+    const std::vector<std::string> devices = ie.GetAvailableDevices();
+    for (std::vector<std::string>::const_iterator i = devices.begin(); i != devices.end(); ++i)
+    {
+        if (i->find("MYRIAD") != std::string::npos)
+        {
+            const std::string name = ie.GetMetric(*i, METRIC_KEY(FULL_DEVICE_NAME)).as<std::string>();
+            CV_LOG_INFO(NULL, "Myriad device: " << name);
+            return name.find("MyriadX") != std::string::npos  || name.find("Myriad X") != std::string::npos;
+        }
+    }
+    return false;
+#else
     InferenceEngine::Builder::Network builder("");
     InferenceEngine::idx_t inpId = builder.addLayer(
                                    InferenceEngine::Builder::InputLayer().setPort(InferenceEngine::Port({1})));
@@ -634,6 +683,7 @@ static bool detectMyriadX_()
         return false;
     }
     return true;
+#endif
 }
 #endif  // !defined(OPENCV_DNN_IE_VPU_TYPE_DEFAULT)
 
@@ -723,7 +773,14 @@ void InfEngineBackendNet::initPlugin(InferenceEngine::CNNNetwork& net)
             // Some of networks can work without a library of extra layers.
 #if INF_ENGINE_VER_MAJOR_GT(INF_ENGINE_RELEASE_2019R1)
             // OpenCV fallbacks as extensions.
-            ie.AddExtension(std::make_shared<InfEngineExtension>(), "CPU");
+            try
+            {
+                ie.AddExtension(std::make_shared<InfEngineExtension>(), "CPU");
+            }
+            catch(const std::exception& e)
+            {
+                CV_LOG_INFO(NULL, "DNN-IE: Can't register OpenCV custom layers extension: " << e.what());
+            }
 #endif
 #ifndef _WIN32
             // Limit the number of CPU threads.
@@ -1052,8 +1109,14 @@ void resetMyriadDevice()
 #if INF_ENGINE_VER_MAJOR_LE(INF_ENGINE_RELEASE_2019R1)
     getSharedPlugins().erase("MYRIAD");
 #else
-    // To unregister both "MYRIAD" and "HETERO:MYRIAD,CPU" plugins
-    getCore() = InferenceEngine::Core();
+    // Unregister both "MYRIAD" and "HETERO:MYRIAD,CPU" plugins
+    InferenceEngine::Core& ie = getCore();
+    try
+    {
+        ie.UnregisterPlugin("MYRIAD");
+        ie.UnregisterPlugin("HETERO");
+    }
+    catch (...) {}
 #endif
 #endif  // HAVE_INF_ENGINE
 }

modules/dnn/src/op_inf_engine.hpp

@@ -23,10 +23,11 @@
 #define INF_ENGINE_RELEASE_2019R1 2019010000
 #define INF_ENGINE_RELEASE_2019R2 2019020000
 #define INF_ENGINE_RELEASE_2019R3 2019030000
+#define INF_ENGINE_RELEASE_2020_1 2020010000
 
 #ifndef INF_ENGINE_RELEASE
-#warning("IE version have not been provided via command-line. Using 2019R3 by default")
-#define INF_ENGINE_RELEASE INF_ENGINE_RELEASE_2019R3
+#warning("IE version have not been provided via command-line. Using 2019.1 by default")
+#define INF_ENGINE_RELEASE INF_ENGINE_RELEASE_2020_1
 #endif
 
 #define INF_ENGINE_VER_MAJOR_GT(ver) (((INF_ENGINE_RELEASE) / 10000) > ((ver) / 10000))

modules/dnn/src/tensorflow/tf_graph_simplifier.cpp

@@ -9,6 +9,7 @@
 
 #ifdef HAVE_PROTOBUF
 
+#include "../graph_simplifier.hpp"
 #include "tf_graph_simplifier.hpp"
 #include <queue>
 
@@ -18,203 +19,87 @@ CV__DNN_INLINE_NS_BEGIN
 using ::google::protobuf::RepeatedField;
 using ::google::protobuf::MapPair;
 
-class Subgraph  // Interface to match and replace TensorFlow subgraphs.
+class TFNodeWrapper : public ImportNodeWrapper
 {
 public:
-    virtual ~Subgraph() {}
+    TFNodeWrapper(tensorflow::NodeDef* _node) : node(_node) {}
 
-    // Add a node to be matched in the origin graph. Specify ids of nodes that
-    // are expected to be inputs. Returns id of a newly added node.
-    // TODO: Replace inputs to std::vector<int> in C++11
-    int addNodeToMatch(const std::string& op, int input_0 = -1, int input_1 = -1,
-                       int input_2 = -1, int input_3 = -1)
+    virtual int getNumInputs() const CV_OVERRIDE
     {
-        int nodeInputs[] = {input_0, input_1, input_2, input_3};
-        int numInputs = 0;
-        for (int i = 0; i < 4; ++i)
-        {
-            numInputs += (int)(nodeInputs[i] != -1);
-        }
-        return addNodeToMatch(op, std::vector<int>(&nodeInputs[0], &nodeInputs[0] + numInputs));
+        return node->input_size();
     }
 
-    int addNodeToMatch(const std::string& op, const std::vector<int>& inputs_)
+    virtual std::string getInputName(int idx) const CV_OVERRIDE
     {
-        for (int i = 0; i < inputs_.size(); ++i)
-        {
-            CV_Assert(inputs_[i] < (int)nodes.size());
-        }
-        nodes.push_back(op);
-        inputs.push_back(inputs_);
-        return nodes.size() - 1;
+        return node->input(idx);
     }
 
-    // Specify resulting node. All the matched nodes in subgraph excluding
-    // input nodes will be fused into this single node.
-    // TODO: Replace inputs to std::vector<int> in C++11
-    void setFusedNode(const std::string& op, int input_0 = -1, int input_1 = -1,
-                      int input_2 = -1, int input_3 = -1, int input_4 = -1,
-                      int input_5 = -1)
+    virtual std::string getType() const CV_OVERRIDE
     {
-        int nodeInputs[] = {input_0, input_1, input_2, input_3, input_4, input_5};
-        int numInputs = 0;
-        for (int i = 0; i < 6; ++i)
-        {
-            CV_Assert(nodeInputs[i] < (int)nodes.size());
-            numInputs += (int)(nodeInputs[i] != -1);
-        }
-        setFusedNode(op, std::vector<int>(&nodeInputs[0], &nodeInputs[0] + numInputs));
+        return node->op();
     }
 
-    void setFusedNode(const std::string& op, const std::vector<int>& inputs_)
+    virtual void setType(const std::string& type) CV_OVERRIDE
     {
-        fusedNodeInputs = inputs_;
-        fusedNodeOp = op;
+        node->set_op(type);
     }
 
-    static int getInputNodeId(const tensorflow::GraphDef& net,
-                              const tensorflow::NodeDef& node,
-                              int inpId)
+    virtual void setInputNames(const std::vector<std::string>& inputs) CV_OVERRIDE
     {
-        CV_Assert(inpId < node.input_size());
-        std::string name = node.input(inpId);
-        // If operation produces several tensors, they are specified by index
-        // after ':' character. In example, "input:0".
-        name = name.substr(0, name.rfind(':'));
-        const int numNodes = net.node_size();
-        for (int i = 0; i < numNodes; ++i)
-        {
-            if (net.node(i).name() == name)
-                return i;
-        }
-        CV_Error(Error::StsParseError, "Input node with name " + name + " not found");
-    }
-
-    // Match TensorFlow subgraph starting from <nodeId> with a set of nodes to be fused.
-    // Const nodes are skipped during matching. Returns true if nodes are matched and can be fused.
-    virtual bool match(const tensorflow::GraphDef& net, int nodeId,
-                       std::vector<int>& matchedNodesIds,
-                       std::vector<int>& targetNodesIds)
-    {
-        matchedNodesIds.clear();
-        targetNodesIds.clear();
-
-        std::queue<int> nodesToMatch;
-        std::queue<int> targetNodes;
-        nodesToMatch.push(nodeId);
-        targetNodes.push(nodes.size() - 1);
-        while (!nodesToMatch.empty())
-        {
-            int nodeToMatch = nodesToMatch.front();
-            int targetNodeId = targetNodes.front();
-            nodesToMatch.pop();
-            targetNodes.pop();
-
-            if (std::find(matchedNodesIds.begin(), matchedNodesIds.end(), nodeToMatch) !=
-                matchedNodesIds.end())
-                continue;
-
-            const tensorflow::NodeDef& node = net.node(nodeToMatch);
-            if (node.op() != nodes[targetNodeId])
-                return false;
-
-            std::vector<int>& inputNodes = inputs[targetNodeId];
-            if (inputNodes.size() != node.input_size())
-                return false;
-
-            for (int j = 0; j < inputNodes.size(); ++j)
-            {
-                if (nodes[inputNodes[j]].empty())  // Unknown input node type.
-                    continue;
-                nodeId = getInputNodeId(net, node, j);
-                const tensorflow::NodeDef& inpNode = net.node(nodeId);
-                if (inpNode.op() != "Const")
-                {
-                    nodesToMatch.push(nodeId);
-                    targetNodes.push(inputNodes[j]);
-                }
-                else if (nodes[inputNodes[j]] != "Const")
-                    return false;
-            }
-            matchedNodesIds.push_back(nodeToMatch);
-            targetNodesIds.push_back(targetNodeId);
-        }
-
-        const int n = matchedNodesIds.size();
-        std::vector<std::pair<int, int> > elements(n);
-        for (int i = 0; i < n; ++i)
-            elements[i] = std::make_pair(matchedNodesIds[i], targetNodesIds[i]);
-        std::sort(elements.begin(), elements.end());
-        for (int i = 0; i < n; ++i)
-        {
-            matchedNodesIds[i] = elements[i].first;
-            targetNodesIds[i] = elements[i].second;
-        }
-        return true;
-    }
-
-    // Fuse matched subgraph.
-    void replace(tensorflow::GraphDef& net, const std::vector<int>& matchedNodesIds,
-                 const std::vector<int>& targetNodesIds)
-    {
-        // Extract names of input nodes.
-        std::vector<std::string> inputsNames(fusedNodeInputs.size());
-        for (int i = 0; i < fusedNodeInputs.size(); ++i)
-        {
-            std::string inpName;
-            // Find input node name looking at inputs of fused nodes.
-            for (int j = 0; j < matchedNodesIds.size() && inpName.empty(); ++j)
-            {
-                const tensorflow::NodeDef &node = net.node(matchedNodesIds[j]);
-                std::vector<int>& inpIndices = inputs[targetNodesIds[j]];
-
-                CV_Assert(node.input_size() == inpIndices.size());
-                for (int k = 0; k < inpIndices.size(); ++k)
-                {
-                    if (inpIndices[k] == fusedNodeInputs[i])
-                    {
-                        inpName = node.input(k);
-                        break;
-                    }
-                }
-            }
-            CV_Assert(!inpName.empty());
-            inputsNames[i] = inpName;
-        }
-
-        // Remove matched nodes except the last one. Indices in ascending order are expected.
-        tensorflow::NodeDef* node = net.mutable_node(matchedNodesIds.back());
-        for (int i = matchedNodesIds.size() - 2; i >= 0; --i)
-            net.mutable_node()->DeleteSubrange(matchedNodesIds[i], 1);
-
-        // Modify the last node to be a fused one.
-        node->set_op(fusedNodeOp);
         node->clear_input();
-        for (int i = 0; i < inputsNames.size(); ++i)
-        {
-            node->add_input(inputsNames[i]);
-        }
-
-        std::vector<tensorflow::NodeDef*> inputNodes(inputsNames.size());
-        for (int i = 0; i < inputsNames.size(); ++i)
-        {
-            inputNodes[i] = net.mutable_node(getInputNodeId(net, *node, i));
-        }
-        finalize(net, node, inputNodes);
+        for (int i = 0; i < inputs.size(); ++i)
+            node->add_input(inputs[i]);
     }
 
-    virtual void finalize(tensorflow::GraphDef&, tensorflow::NodeDef*,
-                          std::vector<tensorflow::NodeDef*>&) {}
-
-private:
-    std::vector<std::string> nodes;         // Nodes to be matched in the origin graph.
-    std::vector<std::vector<int> > inputs;  // Connections of an every node to it's inputs.
-
-    std::string fusedNodeOp;           // Operation name of resulting fused node.
-    std::vector<int> fusedNodeInputs;  // Inputs of fused node.
+    tensorflow::NodeDef* node;
 };
 
-class BatchNormSubgraph : public Subgraph
+class TFGraphWrapper : public ImportGraphWrapper
+{
+public:
+    TFGraphWrapper(tensorflow::GraphDef& _net) : net(_net) {}
+
+    virtual Ptr<ImportNodeWrapper> getNode(int idx) const CV_OVERRIDE
+    {
+        return makePtr<TFNodeWrapper>(net.mutable_node(idx));
+    }
+
+    virtual int getNumNodes() const CV_OVERRIDE
+    {
+        return net.node_size();
+    }
+
+    virtual std::string getNodeName(int idx) const CV_OVERRIDE
+    {
+        return net.node(idx).name();
+    }
+
+    virtual void removeNode(int idx) CV_OVERRIDE
+    {
+        net.mutable_node()->DeleteSubrange(idx, 1);
+    }
+
+    tensorflow::GraphDef& net;
+};
+
+class TFSubgraph : public Subgraph
+{
+    virtual void finalize(const Ptr<ImportGraphWrapper>& netWrapper,
+                          const Ptr<ImportNodeWrapper>& fusedNodeWrapper,
+                          std::vector<Ptr<ImportNodeWrapper> >& inputs) CV_OVERRIDE
+    {
+        std::vector<tensorflow::NodeDef*> inputNodes(inputs.size());
+        for (int i = 0; i < inputs.size(); ++i)
+            inputNodes[i] = inputs[i].dynamicCast<TFNodeWrapper>()->node;
+        finalize(netWrapper.dynamicCast<TFGraphWrapper>()->net,
+                 fusedNodeWrapper.dynamicCast<TFNodeWrapper>()->node, inputNodes);
+    }
+
+    virtual void finalize(tensorflow::GraphDef&, tensorflow::NodeDef* fusedNode,
+                          std::vector<tensorflow::NodeDef*>& inputNodes) {}
+};
+
+class BatchNormSubgraph : public TFSubgraph
 {
 public:
     BatchNormSubgraph()
@@ -250,7 +135,7 @@ public:
     }
 };
 
-class BatchNormNoGammaSubgraph : public Subgraph
+class BatchNormNoGammaSubgraph : public TFSubgraph
 {
 public:
     BatchNormNoGammaSubgraph()
@@ -366,20 +251,21 @@ public:
         setFusedNode("Relu6", input);
     }
 
-    virtual bool match(const tensorflow::GraphDef& net, int nodeId,
+    virtual bool match(const Ptr<ImportGraphWrapper>& net, int nodeId,
                        std::vector<int>& matchedNodesIds,
                        std::vector<int>& targetNodesIds) CV_OVERRIDE
     {
         if (!Subgraph::match(net, nodeId, matchedNodesIds, targetNodesIds))
             return false;
-        Mat maxValue = getTensorContent(net.node(matchedNodesIds.front() + 1).attr().at("value").tensor());
+        tensorflow::NodeDef* node = net->getNode(matchedNodesIds.front() + 1).dynamicCast<TFNodeWrapper>()->node;
+        Mat maxValue = getTensorContent(node->attr().at("value").tensor());
         return maxValue.type() == CV_32FC1 && maxValue.total() == 1 && maxValue.at<float>(0) == 6;
     }
 };
 
 // Keras' reshape stores output shape in separate Const nodes by one value.
 // Need to merge them into a single Const node.
-class ReshapeKerasSubgraph : public Subgraph
+class ReshapeKerasSubgraph : public TFSubgraph
 {
 public:
     ReshapeKerasSubgraph(int _numOutDims) : numOutDims(_numOutDims)
@@ -402,15 +288,15 @@ public:
         setFusedNode("Reshape", ids);
     }
 
-    virtual bool match(const tensorflow::GraphDef& net, int nodeId,
+    virtual bool match(const Ptr<ImportGraphWrapper>& net, int nodeId,
                        std::vector<int>& matchedNodesIds,
                        std::vector<int>& targetNodesIds) CV_OVERRIDE
     {
-        const tensorflow::NodeDef& node = net.node(nodeId);
-        if (node.input_size() == 0)
+        Ptr<ImportNodeWrapper> node = net->getNode(nodeId);
+        if (node->getNumInputs() == 0)
             return false;
 
-        inpName = node.input(0);
+        inpName = node->getInputName(0);
         return Subgraph::match(net, nodeId, matchedNodesIds, targetNodesIds);
     }
 
@@ -457,7 +343,7 @@ public:
     }
 };
 
-class DeconvolutionValidKerasSubgraph : public Subgraph
+class DeconvolutionValidKerasSubgraph : public TFSubgraph
 {
 public:
     DeconvolutionValidKerasSubgraph()
@@ -518,7 +404,7 @@ public:
     }
 };
 
-class DeconvolutionSameKerasSubgraph : public Subgraph
+class DeconvolutionSameKerasSubgraph : public TFSubgraph
 {
 public:
     DeconvolutionSameKerasSubgraph()
@@ -608,7 +494,7 @@ public:
 };
 
 // In case of resizing by factor.
-class UpsamplingKerasSubgraph : public Subgraph
+class UpsamplingKerasSubgraph : public TFSubgraph
 {
 public:
     UpsamplingKerasSubgraph(const std::string& type)
@@ -703,7 +589,7 @@ public:
     }
 };
 
-class KerasMVNSubgraph : public Subgraph
+class KerasMVNSubgraph : public TFSubgraph
 {
 public:
     KerasMVNSubgraph()
@@ -758,20 +644,7 @@ void simplifySubgraphs(tensorflow::GraphDef& net)
     subgraphs.push_back(Ptr<Subgraph>(new ReshapeAsShapeSubgraph()));
     subgraphs.push_back(Ptr<Subgraph>(new KerasMVNSubgraph()));
 
-    int numNodes = net.node_size();
-    std::vector<int> matchedNodesIds, targetNodesIds;
-    for (int i = 0; i < numNodes; ++i)
-    {
-        for (int j = 0; j < subgraphs.size(); ++j)
-        {
-            if (subgraphs[j]->match(net, i, matchedNodesIds, targetNodesIds))
-            {
-                subgraphs[j]->replace(net, matchedNodesIds, targetNodesIds);
-                numNodes -= matchedNodesIds.size() - 1;  // #matchedNodes removed and one added.
-                break;
-            }
-        }
-    }
+    simplifySubgraphs(Ptr<ImportGraphWrapper>(new TFGraphWrapper(net)), subgraphs);
 }
 
 void RemoveIdentityOps(tensorflow::GraphDef& net)
@@ -950,6 +823,7 @@ void sortByExecutionOrder(tensorflow::GraphDef& net)
     for (int i = 0; i < net.node_size(); ++i)
     {
         const tensorflow::NodeDef& node = net.node(i);
+        int numInputsInGraph = 0;
         for (int j = 0; j < node.input_size(); ++j)
         {
             std::string inpName = node.input(j);
@@ -957,22 +831,25 @@ void sortByExecutionOrder(tensorflow::GraphDef& net)
             inpName = inpName.substr(inpName.find('^') + 1);
 
             nodesMapIt = nodesMap.find(inpName);
-            CV_Assert(nodesMapIt != nodesMap.end());
-            edges[nodesMapIt->second].push_back(i);
+            if (nodesMapIt != nodesMap.end())
+            {
+                edges[nodesMapIt->second].push_back(i);
+                numInputsInGraph += 1;
+            }
         }
-        if (node.input_size() == 0)
+        if (numInputsInGraph == 0)
             nodesToAdd.push_back(i);
         else
         {
             if (node.op() == "Merge" || node.op() == "RefMerge")
             {
                 int numControlEdges = 0;
-                for (int j = 0; j < node.input_size(); ++j)
+                for (int j = 0; j < numInputsInGraph; ++j)
                     numControlEdges += node.input(j)[0] == '^';
                 numRefsToAdd[i] = numControlEdges + 1;
             }
             else
-                numRefsToAdd[i] = node.input_size();
+                numRefsToAdd[i] = numInputsInGraph;
         }
     }
 

modules/dnn/src/tensorflow/tf_importer.cpp

@@ -715,6 +715,10 @@ void TFImporter::populateNet(Net dstNet)
         simplifySubgraphs(netBin);
         sortByExecutionOrder(netBin);
     }
+    else
+    {
+        sortByExecutionOrder(netTxt);
+    }
 
     std::set<String> layers_to_ignore;
 
@@ -996,7 +1000,7 @@ void TFImporter::populateNet(Net dstNet)
             if (getDataLayout(name, data_layouts) == DATA_LAYOUT_UNKNOWN)
                 data_layouts[name] = DATA_LAYOUT_NHWC;
         }
-        else if (type == "BiasAdd" || type == "Add" || type == "Sub" || type=="AddN")
+        else if (type == "BiasAdd" || type == "Add" || type == "AddV2" || type == "Sub" || type=="AddN")
         {
             bool haveConst = false;
             for(int ii = 0; !haveConst && ii < layer.input_size(); ++ii)

modules/dnn/test/test_backends.cpp

@@ -197,8 +197,8 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_Caffe_Different_Width_Height)
     if (backend == DNN_BACKEND_HALIDE)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_HALIDE);
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) &&
+        target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);
 #endif
     Mat sample = imread(findDataFile("dnn/street.png"));
@@ -249,8 +249,8 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_v1_TensorFlow_Different_Width_Height)
     if (backend == DNN_BACKEND_HALIDE)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_HALIDE);
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) &&
+        target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);
 #endif
 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2019020000)

modules/dnn/test/test_caffe_importer.cpp

@@ -691,9 +691,11 @@ TEST_P(Test_Caffe_nets, FasterRCNN_zf)
         (target == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_512MB : CV_TEST_TAG_MEMORY_1GB),
         CV_TEST_TAG_DEBUG_LONG
     );
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_OPENCL_FP16)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && target == DNN_TARGET_OPENCL_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD);
     if (target == DNN_TARGET_CUDA_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
@@ -710,9 +712,11 @@ TEST_P(Test_Caffe_nets, RFCN)
         CV_TEST_TAG_LONG,
         CV_TEST_TAG_DEBUG_VERYLONG
     );
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_OPENCL_FP16)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && target == DNN_TARGET_OPENCL_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD);
     float scoreDiff = default_l1, iouDiff = default_lInf;
     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)

modules/dnn/test/test_darknet_importer.cpp

@@ -307,8 +307,8 @@ TEST_P(Test_Darknet_nets, YoloVoc)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
 #endif
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) &&
+        target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);  // need to update check function
 #endif
 
@@ -352,8 +352,8 @@ TEST_P(Test_Darknet_nets, TinyYoloVoc)
     applyTestTag(CV_TEST_TAG_MEMORY_512MB);
 
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 || backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) &&
+        target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);  // need to update check function
 #endif
     // batchId, classId, confidence, left, top, right, bottom
@@ -486,7 +486,8 @@ TEST_P(Test_Darknet_nets, YOLOv3)
     std::string weights_file = "yolov3.weights";
 
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD &&
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && target == DNN_TARGET_MYRIAD &&
         getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
     {
         scoreDiff = 0.04;

modules/dnn/test/test_halide_layers.cpp

@@ -357,11 +357,6 @@ TEST_P(MaxPooling, Accuracy)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
 #endif
 
-#if defined(INF_ENGINE_RELEASE)
-    if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && stride != Size(1, 1) && pad != Size(0, 0))
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
-#endif
-
     LayerParams lp;
     lp.set("pool", "max");
     lp.set("kernel_w", kernel.width);
@@ -399,7 +394,8 @@ TEST_P(FullyConnected, Accuracy)
     bool hasBias = get<3>(GetParam());
     Backend backendId = get<0>(get<4>(GetParam()));
     Target targetId = get<1>(get<4>(GetParam()));
-    if (backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && (targetId == DNN_TARGET_OPENCL_FP16 ||
+    if ((backendId == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backendId == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && (targetId == DNN_TARGET_OPENCL_FP16 ||
        (targetId == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X))) {
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);

modules/dnn/test/test_ie_models.cpp

@@ -134,12 +134,13 @@ static const std::vector<std::string> getOpenVINOTestModelsList()
     return result;
 }
 
-static inline void genData(const std::vector<size_t>& dims, Mat& m, Blob::Ptr& dataPtr)
+static inline void genData(const InferenceEngine::TensorDesc& desc, Mat& m, Blob::Ptr& dataPtr)
 {
+    const std::vector<size_t>& dims = desc.getDims();
     m.create(std::vector<int>(dims.begin(), dims.end()), CV_32F);
     randu(m, -1, 1);
 
-    dataPtr = make_shared_blob<float>({Precision::FP32, dims, Layout::ANY}, (float*)m.data);
+    dataPtr = make_shared_blob<float>(desc, (float*)m.data);
 }
 
 void runIE(Target target, const std::string& xmlPath, const std::string& binPath,
@@ -238,7 +239,7 @@ void runIE(Target target, const std::string& xmlPath, const std::string& binPath
     BlobMap inputBlobs;
     for (auto& it : net.getInputsInfo())
     {
-        genData(it.second->getTensorDesc().getDims(), inputsMap[it.first], inputBlobs[it.first]);
+        genData(it.second->getTensorDesc(), inputsMap[it.first], inputBlobs[it.first]);
     }
     infRequest.SetInput(inputBlobs);
 
@@ -247,7 +248,7 @@ void runIE(Target target, const std::string& xmlPath, const std::string& binPath
     BlobMap outputBlobs;
     for (auto& it : net.getOutputsInfo())
     {
-        genData(it.second->getTensorDesc().getDims(), outputsMap[it.first], outputBlobs[it.first]);
+        genData(it.second->getTensorDesc(), outputsMap[it.first], outputBlobs[it.first]);
     }
     infRequest.SetOutput(outputBlobs);
 

modules/dnn/test/test_layers.cpp

@@ -864,6 +864,8 @@ TEST_P(Test_Caffe_layers, PriorBox_squares)
 {
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_MYRIAD)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
     LayerParams lp;
     lp.name = "testPriorBox";
     lp.type = "PriorBox";
@@ -1301,7 +1303,7 @@ static void test_dldt_fused_output(Backend backend, Target target)
     }
     net.setPreferableBackend(backend);
     net.setPreferableTarget(target);
-    net.setInput(Mat({1, 1, 1, 1}, CV_32FC1, Scalar(1)));
+    net.setInput(Mat({1, 1, 2, 3}, CV_32FC1, Scalar(1)));
     net.forward();
 }
 
@@ -1340,7 +1342,7 @@ TEST_P(Test_DLDT_layers, multiple_networks)
         nets[i].addLayerToPrev(lp.name, lp.type, lp);
         nets[i].setPreferableBackend(backend);
         nets[i].setPreferableTarget(target);
-        nets[i].setInput(Mat({1, 1, 1, 1}, CV_32FC1, Scalar(1)));
+        nets[i].setInput(Mat({1, 1, 2, 3}, CV_32FC1, Scalar(1)));
     }
     Mat out_1 = nets[0].forward();
     Mat out_2 = nets[1].forward();

modules/dnn/test/test_onnx_importer.cpp

@@ -369,9 +369,12 @@ TEST_P(Test_ONNX_layers, Div)
     net.setPreferableBackend(backend);
     net.setPreferableTarget(target);
 
-    Mat inp1 = blobFromNPY(_tf("data/input_div_0.npy"));
-    Mat inp2 = blobFromNPY(_tf("data/input_div_1.npy"));
+    // Reference output values range is -68.80928, 2.991873. So to avoid computational
+    // difference for FP16 we'll perform reversed division (just swap inputs).
+    Mat inp1 = blobFromNPY(_tf("data/input_div_1.npy"));
+    Mat inp2 = blobFromNPY(_tf("data/input_div_0.npy"));
     Mat ref  = blobFromNPY(_tf("data/output_div.npy"));
+    cv::divide(1.0, ref, ref);
     checkBackend(&inp1, &ref);
 
     net.setInput(inp1, "0");
@@ -421,6 +424,7 @@ TEST_P(Test_ONNX_layers, Softmax)
 {
     testONNXModels("softmax");
     testONNXModels("log_softmax", npy, 0, 0, false, false);
+    testONNXModels("softmax_unfused");
 }
 
 TEST_P(Test_ONNX_layers, Split_EltwiseMax)
@@ -473,6 +477,9 @@ TEST_P(Test_ONNX_nets, Googlenet)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
 
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+
     const String model = _tf("models/googlenet.onnx", false);
 
     Net net = readNetFromONNX(model);
@@ -516,7 +523,7 @@ TEST_P(Test_ONNX_nets, RCNN_ILSVRC13)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
 #endif
     // Reference output values are in range [-4.992, -1.161]
-    testONNXModels("rcnn_ilsvrc13", pb, 0.0045);
+    testONNXModels("rcnn_ilsvrc13", pb, 0.0046);
 }
 
 TEST_P(Test_ONNX_nets, VGG16_bn)
@@ -583,10 +590,12 @@ TEST_P(Test_ONNX_nets, TinyYolov2)
     )
         applyTestTag(target == DNN_TARGET_OPENCL ? CV_TEST_TAG_DNN_SKIP_IE_OPENCL : CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
 
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
+    if (target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
     )
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X,
+                     backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ?
+                     CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER :
+                     CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
 #endif
 
     // output range: [-11; 8]
@@ -628,6 +637,12 @@ TEST_P(Test_ONNX_nets, LResNet100E_IR)
         if (target == DNN_TARGET_OPENCL)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
         if (target == DNN_TARGET_MYRIAD)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
     }
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
+    {
+        if (target == DNN_TARGET_OPENCL_FP16) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+        if (target == DNN_TARGET_OPENCL)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+        if (target == DNN_TARGET_MYRIAD)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+    }
 
     double l1 = default_l1, lInf = default_lInf;
     // output range: [-3; 3]
@@ -652,10 +667,11 @@ TEST_P(Test_ONNX_nets, LResNet100E_IR)
 TEST_P(Test_ONNX_nets, Emotion_ferplus)
 {
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
-    )
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
+    if (target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X,
+                     backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ?
+                     CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER :
+                     CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
 #endif
 
     double l1 = default_l1;
@@ -692,7 +708,8 @@ TEST_P(Test_ONNX_nets, DenseNet121)
 TEST_P(Test_ONNX_nets, Inception_v1)
 {
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ||
+         backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH) && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD);
 #endif
     testONNXModels("inception_v1", pb);

modules/dnn/test/test_tf_importer.cpp

@@ -261,10 +261,13 @@ TEST_P(Test_TensorFlow_layers, ave_pool_same)
 {
     // Reference output values are in range [-0.519531, 0.112976]
 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_GE(2019010000)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X
-    )
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
+    if (target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    {
+        if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019)
+            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
+        else if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
+            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
+    }
 #endif
     runTensorFlowNet("ave_pool_same");
 }
@@ -399,6 +402,8 @@ TEST_P(Test_TensorFlow_layers, l2_normalize_3d)
 #if defined(INF_ENGINE_RELEASE)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_MYRIAD)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
 #endif
 
     runTensorFlowNet("l2_normalize_3d");
@@ -409,11 +414,15 @@ class Test_TensorFlow_nets : public DNNTestLayer {};
 TEST_P(Test_TensorFlow_nets, MobileNet_SSD)
 {
 #if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
+    if (target == DNN_TARGET_MYRIAD)
     {
 #if INF_ENGINE_VER_MAJOR_GE(2019020000)
         if (getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
-            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
+            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X,
+                         backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ?
+                             CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER :
+                             CV_TEST_TAG_DNN_SKIP_IE_NGRAPH,
+                         CV_TEST_TAG_DNN_SKIP_IE_VERSION);
 #endif
     }
 #endif
@@ -554,6 +563,10 @@ TEST_P(Test_TensorFlow_nets, Faster_RCNN)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 &&
         (INF_ENGINE_VER_MAJOR_LT(2019020000) || target != DNN_TARGET_CPU))
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER, CV_TEST_TAG_DNN_SKIP_IE_VERSION);
+
+    if (INF_ENGINE_VER_MAJOR_GT(2019030000) &&
+        backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_MYRIAD)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
 #endif
     // segfault: inference-engine/thirdparty/clDNN/src/gpu/detection_output_cpu.cpp:111:
     // Assertion `prior_height > 0' failed.

modules/dnn/test/test_torch_importer.cpp

@@ -239,6 +239,8 @@ TEST_P(Test_Torch_layers, net_conv_gemm_lrn)
 {
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_MYRIAD)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
     double l1 = 0.0, lInf = 0.0;
     if (target == DNN_TARGET_OPENCL_FP16)
     {
@@ -398,6 +400,13 @@ TEST_P(Test_Torch_nets, ENet_accuracy)
         if (target == DNN_TARGET_MYRIAD)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
         throw SkipTestException("");
     }
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target != DNN_TARGET_CPU)
+    {
+        if (target == DNN_TARGET_OPENCL_FP16) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+        if (target == DNN_TARGET_OPENCL)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+        if (target == DNN_TARGET_MYRIAD)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
+        throw SkipTestException("");
+    }
 
     Net net;
     {
@@ -450,6 +459,9 @@ TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD
             && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH && target == DNN_TARGET_MYRIAD
+            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X, CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
 #endif
 
     checkBackend();

samples/dnn/tf_text_graph_ssd.py

@@ -62,7 +62,7 @@ class MultiscaleAnchorGenerator:
 
 def createSSDGraph(modelPath, configPath, outputPath):
     # Nodes that should be kept.
-    keepOps = ['Conv2D', 'BiasAdd', 'Add', 'Relu', 'Relu6', 'Placeholder', 'FusedBatchNorm',
+    keepOps = ['Conv2D', 'BiasAdd', 'Add', 'AddV2', 'Relu', 'Relu6', 'Placeholder', 'FusedBatchNorm',
                'DepthwiseConv2dNative', 'ConcatV2', 'Mul', 'MaxPool', 'AvgPool', 'Identity',
                'Sub', 'ResizeNearestNeighbor', 'Pad', 'FusedBatchNormV3']
 
@@ -151,6 +151,9 @@ def createSSDGraph(modelPath, configPath, outputPath):
         subgraphBatchNorm = ['Add',
             ['Mul', 'input', ['Mul', ['Rsqrt', ['Add', 'moving_variance', 'add_y']], 'gamma']],
             ['Sub', 'beta', ['Mul', 'moving_mean', 'Mul_0']]]
+        subgraphBatchNormV2 = ['AddV2',
+            ['Mul', 'input', ['Mul', ['Rsqrt', ['AddV2', 'moving_variance', 'add_y']], 'gamma']],
+            ['Sub', 'beta', ['Mul', 'moving_mean', 'Mul_0']]]
         # Detect unfused nearest neighbor resize.
         subgraphResizeNN = ['Reshape',
             ['Mul', ['Reshape', 'input', ['Pack', 'shape_1', 'shape_2', 'shape_3', 'shape_4', 'shape_5']],
@@ -177,7 +180,8 @@ def createSSDGraph(modelPath, configPath, outputPath):
         for node in graph_def.node:
             inputs = {}
             fusedNodes = []
-            if checkSubgraph(node, subgraphBatchNorm, inputs, fusedNodes):
+            if checkSubgraph(node, subgraphBatchNorm, inputs, fusedNodes) or \
+               checkSubgraph(node, subgraphBatchNormV2, inputs, fusedNodes):
                 name = node.name
                 node.Clear()
                 node.name = name