dnn(ie): fix handling of 1D and non-32F outputs of InferenceEngine

modules/dnn/src/dnn.cpp

@@ -1944,7 +1944,10 @@ struct Net::Impl : public detail::NetImplBase
 
             Ptr<InfEngineNgraphNode> ieNode = node.dynamicCast<InfEngineNgraphNode>();
             CV_Assert(!ieNode.empty());
-            ieNode->net->reset();
+
+            CV_Assert(ieNode->net);
+            InfEngineNgraphNet& ienet = *ieNode->net;
+            ienet.reset();
 
             for (it = layers.begin(); it != layers.end(); ++it)
             {
@@ -1961,16 +1964,26 @@ struct Net::Impl : public detail::NetImplBase
                 {
                     for (int i = 0; i < ld.outputBlobsWrappers.size(); ++i)
                     {
-                        InferenceEngine::DataPtr dataPtr = ngraphDataNode(ld.outputBlobsWrappers[i]);
-                        dataPtr->setName(ld.name);
+                        auto it = ienet.outputsDesc.find(ld.name);
+                        if (it != ienet.outputsDesc.end())
+                        {
+                            const InferenceEngine::TensorDesc& descriptor = it->second;
+                            InferenceEngine::DataPtr dataPtr = ngraphDataOutputNode(ld.outputBlobsWrappers[i], descriptor, ld.name);
+                            dataPtr->setName(ld.name);
+                        }
+                        else
+                        {
+                            InferenceEngine::DataPtr dataPtr = ngraphDataNode(ld.outputBlobsWrappers[i]);
+                            dataPtr->setName(ld.name);
+                        }
                     }
                 }
-                ieNode->net->addBlobs(ld.inputBlobsWrappers);
-                ieNode->net->addBlobs(ld.outputBlobsWrappers);
+                ienet.addBlobs(ld.inputBlobsWrappers);
+                ienet.addBlobs(ld.outputBlobsWrappers);
                 ld.skip = true;
             }
             layers[lastLayerId].skip = false;
-            ieNode->net->init((Target)preferableTarget);
+            ienet.init((Target)preferableTarget);
             return;
         }
 
@@ -3719,8 +3732,8 @@ void Net::forward(OutputArrayOfArrays outputBlobs,
         matvec.push_back(impl->getBlob(pins[i]));
     }
 
-    std::vector<Mat> & outputvec = *(std::vector<Mat> *)outputBlobs.getObj();
-    outputvec = matvec;
+    outputBlobs.create((int)matvec.size(), 1, CV_32F/*FIXIT*/, -1);  // allocate vector
+    outputBlobs.assign(matvec);
 }
 
 void Net::forward(std::vector<std::vector<Mat> >& outputBlobs,

modules/dnn/src/ie_ngraph.cpp

@@ -789,21 +789,32 @@ void NgraphBackendLayer::forward(InputArrayOfArrays inputs, OutputArrayOfArrays
 }
 
 
-static InferenceEngine::Layout estimateLayout(const Mat& m)
+static InferenceEngine::Layout estimateLayout(int dims)
 {
-    if (m.dims == 4)
+    if (dims == 4)
         return InferenceEngine::Layout::NCHW;
-    else if (m.dims == 3)
+    else if (dims == 3)
         return InferenceEngine::Layout::CHW;
-    else if (m.dims == 2)
+    else if (dims == 2)
         return InferenceEngine::Layout::NC;
-    else if (m.dims == 1)
+    else if (dims == 1)
         return InferenceEngine::Layout::C;
-    else if (m.dims == 5)
+    else if (dims == 5)
         return InferenceEngine::Layout::NCDHW;
     else
         return InferenceEngine::Layout::ANY;
 }
+static inline
+InferenceEngine::Layout estimateLayout(size_t dims)
+{
+    return estimateLayout((int)dims);
+}
+
+static inline
+InferenceEngine::Layout estimateLayout(const Mat& m)
+{
+    return estimateLayout(m.dims);
+}
 
 static InferenceEngine::DataPtr wrapToInfEngineDataNode(const Mat& m, const std::string& name = "")
 {
@@ -839,6 +850,7 @@ InferenceEngine::Blob::Ptr wrapToNgraphBlob(const Mat& m, InferenceEngine::Layou
 
 NgraphBackendWrapper::NgraphBackendWrapper(int targetId, const cv::Mat& m)
     : BackendWrapper(DNN_BACKEND_INFERENCE_ENGINE_NGRAPH, targetId)
+    , host((Mat*)&m)
 {
     dataPtr = wrapToInfEngineDataNode(m);
     blob = wrapToNgraphBlob(m, estimateLayout(m));
@@ -890,7 +902,11 @@ InferenceEngine::Blob::Ptr copyBlob(const InferenceEngine::Blob::Ptr& blob)
         copy = InferenceEngine::make_shared_blob<uint8_t>(description);
     }
     else
-        CV_Error(Error::StsNotImplemented, "Unsupported blob precision");
+    {
+        std::ostringstream msg;
+        msg << precision;
+        CV_Error_(Error::StsNotImplemented, ("Unsupported blob precision: %s", msg.str().c_str()));
+    }
     copy->allocate();
     return copy;
 }
@@ -903,6 +919,66 @@ InferenceEngine::DataPtr ngraphDataNode(const Ptr<BackendWrapper>& ptr)
     return p->dataPtr;
 }
 
+static
+InferenceEngine::Blob::Ptr reallocateBlob(Mat &m, const InferenceEngine::TensorDesc& description)
+{
+    auto dims = description.getDims();
+    auto layout = estimateLayout(dims.size());
+    MatShape matShape(dims.begin(), dims.end());
+    if (description.getPrecision() == InferenceEngine::Precision::FP32)
+    {
+        m.create(matShape, CV_32FC1);
+        return InferenceEngine::make_shared_blob<float>(
+                {description.getPrecision(), dims, layout}, (float*)m.data);
+    }
+    else if (description.getPrecision() == InferenceEngine::Precision::I32)
+    {
+        m.create(matShape, CV_32SC1);
+        return InferenceEngine::make_shared_blob<int>(
+                {description.getPrecision(), dims, layout}, (int*)m.data);
+    }
+    else if (description.getPrecision() == InferenceEngine::Precision::U8)
+    {
+        m.create(matShape, CV_8UC1);
+        return InferenceEngine::make_shared_blob<uchar>(
+                {description.getPrecision(), dims, layout}, (uchar*)m.data);
+    }
+    std::ostringstream msg;
+    msg << "Unsupported IE precision: " << description.getPrecision();
+    CV_Error(Error::StsNotImplemented, msg.str());
+}
+
+InferenceEngine::DataPtr ngraphDataOutputNode(
+        const Ptr<BackendWrapper>& ptr,
+        const InferenceEngine::TensorDesc& description,
+        const std::string name)
+{
+    CV_Assert(!ptr.empty());
+    Ptr<NgraphBackendWrapper> p = ptr.dynamicCast<NgraphBackendWrapper>();
+    CV_Assert(!p.empty());
+    NgraphBackendWrapper& w = *p;
+    const InferenceEngine::TensorDesc& blobDesc = w.blob.get()->getTensorDesc();
+    auto dims = description.getDims();
+    bool reallocate = false;
+    if (blobDesc.getPrecision() != description.getPrecision())
+    {
+        reallocate = true;
+        CV_LOG_WARNING(NULL, "Reallocate output '" << name << "' blob due to wrong precision: " << blobDesc.getPrecision() << " => " << description.getPrecision() << "  ndims=" << dims.size());
+    }
+    if (dims.size() != blobDesc.getDims().size())
+    {
+        reallocate = true;
+        CV_LOG_WARNING(NULL, "Reallocate output '" << name << "' blob due to wrong dims: " << blobDesc.getDims().size() << " => " << dims.size());
+    }
+    if (reallocate)
+    {
+        auto layout = estimateLayout(dims.size());
+        w.dataPtr = InferenceEngine::DataPtr(new InferenceEngine::Data(name,
+               {description.getPrecision(), dims, layout}));
+        w.blob = reallocateBlob(*w.host, description);
+    }
+    return w.dataPtr;
+}
 
 void forwardNgraph(const std::vector<Ptr<BackendWrapper> >& outBlobsWrappers,
                       Ptr<BackendNode>& node, bool isAsync)
@@ -918,6 +994,13 @@ void InfEngineNgraphNet::reset()
     allBlobs.clear();
     infRequests.clear();
     isInit = false;
+
+    outputsDesc.clear();
+    for (const auto& it : cnn.getOutputsInfo())
+    {
+        const std::string& name = it.first;
+        outputsDesc.insert({name, it.second->getTensorDesc()});
+    }
 }
 
 void InfEngineNgraphNet::addBlobs(const std::vector<cv::Ptr<BackendWrapper> >& ptrs)

modules/dnn/src/ie_ngraph.hpp

@@ -54,7 +54,8 @@ public:
     void setNodePtr(std::shared_ptr<ngraph::Node>* ptr);
 
     void reset();
-private:
+
+//private:
     detail::NetImplBase& netImpl_;
 
     void release();
@@ -89,6 +90,8 @@ private:
     bool hasNetOwner;
     std::vector<std::string> requestedOutputs;
     std::unordered_set<std::shared_ptr<ngraph::Node>> unconnectedNodes;
+
+    std::map<std::string, InferenceEngine::TensorDesc> outputsDesc;
 };
 
 class InfEngineNgraphNode : public BackendNode
@@ -121,12 +124,17 @@ public:
     virtual void copyToHost() CV_OVERRIDE;
     virtual void setHostDirty() CV_OVERRIDE;
 
+    Mat* host;
     InferenceEngine::DataPtr dataPtr;
     InferenceEngine::Blob::Ptr blob;
     AsyncArray futureMat;
 };
 
 InferenceEngine::DataPtr ngraphDataNode(const Ptr<BackendWrapper>& ptr);
+InferenceEngine::DataPtr ngraphDataOutputNode(
+        const Ptr<BackendWrapper>& ptr,
+        const InferenceEngine::TensorDesc& description,
+        const std::string name);
 
 // This is a fake class to run networks from Model Optimizer. Objects of that
 // class simulate responses of layers are imported by OpenCV and supported by