Merge pull request #13608 from allnes:dnn_rework

modules/dnn/src/dnn.cpp

@@ -2600,7 +2600,7 @@ Net Net::readFromModelOptimizer(const String& xml, const String& bin)
     backendNode->net = Ptr<InfEngineBackendNet>(new InfEngineBackendNet(ieNet));
     for (auto& it : ieNet.getOutputsInfo())
     {
-        Ptr<Layer> cvLayer(new InfEngineBackendLayer(it.second));
+        Ptr<Layer> cvLayer(new InfEngineBackendLayer(ieNet));
         InferenceEngine::CNNLayerPtr ieLayer = ieNet.getLayerByName(it.first.c_str());
         CV_Assert(ieLayer);
 

modules/dnn/src/op_inf_engine.cpp

@@ -718,19 +718,33 @@ Mat infEngineBlobToMat(const InferenceEngine::Blob::Ptr& blob)
     return Mat(size, CV_32F, (void*)blob->buffer());
 }
 
-InfEngineBackendLayer::InfEngineBackendLayer(const InferenceEngine::DataPtr& output_)
-{
-    output = output_;
-}
-
 bool InfEngineBackendLayer::getMemoryShapes(const std::vector<MatShape> &inputs,
                                             const int requiredOutputs,
                                             std::vector<MatShape> &outputs,
                                             std::vector<MatShape> &internals) const
 {
-    std::vector<size_t> dims = output->dims;
+    InferenceEngine::ICNNNetwork::InputShapes inShapes = t_net.getInputShapes();
-    std::vector<int> shape(dims.rbegin(), dims.rend());
+    InferenceEngine::ICNNNetwork::InputShapes::iterator itr;
-    outputs.assign(1, shape);
+    bool equal_flag = true;
+    size_t i = 0;
+    for (itr = inShapes.begin(); itr != inShapes.end(); ++itr)
+    {
+        InferenceEngine::SizeVector currentInShape(inputs[i].begin(), inputs[i].end());
+        if (itr->second != currentInShape)
+        {
+            itr->second = currentInShape;
+            equal_flag = false;
+        }
+        i++;
+    }
+
+    if (!equal_flag)
+    {
+        InferenceEngine::CNNNetwork curr_t_net(t_net);
+        curr_t_net.reshape(inShapes);
+    }
+    std::vector<size_t> dims = t_net.getOutputsInfo()[name]->getDims();
+    outputs.push_back(MatShape(dims.begin(), dims.end()));
     return false;
 }
 

modules/dnn/src/op_inf_engine.hpp

@@ -260,7 +260,7 @@ InferenceEngine::TBlob<int16_t>::Ptr convertFp16(const InferenceEngine::Blob::Pt
 class InfEngineBackendLayer : public Layer
 {
 public:
-    InfEngineBackendLayer(const InferenceEngine::DataPtr& output);
+    InfEngineBackendLayer(const InferenceEngine::CNNNetwork &t_net_) : t_net(t_net_) {};
 
     virtual bool getMemoryShapes(const std::vector<MatShape> &inputs,
                                  const int requiredOutputs,
@@ -273,7 +273,7 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE;
 
 private:
-    InferenceEngine::DataPtr output;
+    InferenceEngine::CNNNetwork t_net;
 };
 
 #endif  // HAVE_INF_ENGINE

modules/dnn/test/test_layers.cpp

@@ -1008,8 +1008,8 @@ INSTANTIATE_TEST_CASE_P(/**/, Layer_Test_Convolution_DLDT,
 // net.save('/path/to/caffemodel')
 //
 // 3. Convert using ModelOptimizer.
-typedef testing::TestWithParam<tuple<int, int, Target> > Test_DLDT_two_inputs;
+typedef testing::TestWithParam<tuple<int, int, Target, std::vector<int> > > Test_DLDT_two_inputs_3dim;
-TEST_P(Test_DLDT_two_inputs, as_IR)
+TEST_P(Test_DLDT_two_inputs_3dim, as_IR)
 {
     int firstInpType = get<0>(GetParam());
     int secondInpType = get<1>(GetParam());
@@ -1021,9 +1021,9 @@ TEST_P(Test_DLDT_two_inputs, as_IR)
 #endif
 
     Net net = readNet(_tf("net_two_inputs.xml"), _tf("net_two_inputs.bin"));
-    int inpSize[] = {1, 2, 3};
+    std::vector<int> inpSize = get<3>(GetParam());
-    Mat firstInp(3, &inpSize[0], firstInpType);
+    Mat firstInp(3, inpSize.data(), firstInpType);
-    Mat secondInp(3, &inpSize[0], secondInpType);
+    Mat secondInp(3, inpSize.data(), secondInpType);
     randu(firstInp, 0, 255);
     randu(secondInp, 0, 255);
 
@@ -1046,6 +1046,15 @@ TEST_P(Test_DLDT_two_inputs, as_IR)
     }
 }
 
+std::vector< std::vector<int> > list_sizes{ {1, 2, 3}, {3, 2, 1}, {5, 5, 5}, {13, 7, 11} };
+
+INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_DLDT_two_inputs_3dim, Combine(
+  Values(CV_8U, CV_32F), Values(CV_8U, CV_32F),
+  testing::ValuesIn(getAvailableTargets(DNN_BACKEND_INFERENCE_ENGINE)),
+  testing::ValuesIn(list_sizes)
+));
+
+typedef testing::TestWithParam<tuple<int, int, Target> > Test_DLDT_two_inputs;
 TEST_P(Test_DLDT_two_inputs, as_backend)
 {
     static const float kScale = 0.5f;