Fix custom IE layers in case of no MKLDNN plugin

modules/dnn/src/dnn.cpp

@@ -97,6 +97,37 @@ public:
         static BackendRegistry impl;
         return impl;
     }
+
+    static inline bool checkIETarget(int target)
+    {
+#ifndef HAVE_INF_ENGINE
+        return false;
+#else
+        cv::dnn::Net net;
+        cv::dnn::LayerParams lp;
+        lp.set("kernel_size", 1);
+        lp.set("num_output", 1);
+        lp.set("bias_term", false);
+        lp.type = "Convolution";
+        lp.name = "testLayer";
+        lp.blobs.push_back(Mat({1, 2, 1, 1}, CV_32F, Scalar(1)));
+        net.addLayerToPrev(lp.name, lp.type, lp);
+        net.setPreferableBackend(cv::dnn::DNN_BACKEND_INFERENCE_ENGINE);
+        net.setPreferableTarget(target);
+        static int inpDims[] = {1, 2, 3, 4};
+        net.setInput(cv::Mat(4, &inpDims[0], CV_32FC1, cv::Scalar(0)));
+        try
+        {
+            net.forward();
+        }
+        catch(...)
+        {
+            return false;
+        }
+        return true;
+#endif
+    }
+
 private:
     BackendRegistry()
     {
@@ -136,35 +167,6 @@ private:
 
         backends.push_back(std::make_pair(DNN_BACKEND_OPENCV, DNN_TARGET_CPU));
     }
-    static inline bool checkIETarget(int target)
-    {
-#ifndef HAVE_INF_ENGINE
-        return false;
-#else
-        cv::dnn::Net net;
-        cv::dnn::LayerParams lp;
-        lp.set("kernel_size", 1);
-        lp.set("num_output", 1);
-        lp.set("bias_term", false);
-        lp.type = "Convolution";
-        lp.name = "testLayer";
-        lp.blobs.push_back(Mat({1, 2, 1, 1}, CV_32F, Scalar(1)));
-        net.addLayerToPrev(lp.name, lp.type, lp);
-        net.setPreferableBackend(cv::dnn::DNN_BACKEND_INFERENCE_ENGINE);
-        net.setPreferableTarget(target);
-        static int inpDims[] = {1, 2, 3, 4};
-        net.setInput(cv::Mat(4, &inpDims[0], CV_32FC1, cv::Scalar(0)));
-        try
-        {
-            net.forward();
-        }
-        catch(...)
-        {
-            return false;
-        }
-        return true;
-#endif
-    }
 
     BackendsList backends;
 };
@@ -1544,6 +1546,9 @@ struct Net::Impl
         // backend. Split a whole model on several Inference Engine networks if
         // some of layers are not implemented.
 
+        bool supportsCPUFallback = preferableTarget == DNN_TARGET_CPU ||
+                                   BackendRegistry::checkIETarget(DNN_TARGET_CPU);
+
         // Set of all input and output blobs wrappers for current network.
         std::map<LayerPin, Ptr<BackendWrapper> > netBlobsWrappers;
         for (it = layers.begin(); it != layers.end(); ++it)
@@ -1557,7 +1562,8 @@ struct Net::Impl
             if (!fused && !layer->supportBackend(preferableBackend))
             {
                 bool customizable = ld.id != 0 && ld.outputBlobs.size() == 1 &&
-                                    INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R2);
+                                    INF_ENGINE_VER_MAJOR_GE(INF_ENGINE_RELEASE_2019R2) &&
+                                    supportsCPUFallback;
                 // TODO: there is a bug in Myriad plugin with custom layers shape infer.
                 if (preferableTarget == DNN_TARGET_MYRIAD)
                 {

modules/dnn/test/test_darknet_importer.cpp

@@ -390,12 +390,6 @@ TEST_P(Test_Darknet_nets, YOLOv3)
 {
     applyTestTag(CV_TEST_TAG_LONG, (target == DNN_TARGET_CPU ? CV_TEST_TAG_MEMORY_1GB : CV_TEST_TAG_MEMORY_2GB));
 
-#if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD
-            && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);
-#endif
-
     // batchId, classId, confidence, left, top, right, bottom
     Mat ref = (Mat_<float>(9, 7) << 0, 7,  0.952983f, 0.614622f, 0.150257f, 0.901369f, 0.289251f,  // a truck
                                     0, 1,  0.987908f, 0.150913f, 0.221933f, 0.742255f, 0.74626f,   // a bicycle
@@ -413,24 +407,36 @@ TEST_P(Test_Darknet_nets, YOLOv3)
     std::string config_file = "yolov3.cfg";
     std::string weights_file = "yolov3.weights";
 
+#if defined(INF_ENGINE_RELEASE)
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD &&
+        getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    {
+        scoreDiff = 0.04;
+        iouDiff = 0.2;
+    }
+#endif
+
     {
     SCOPED_TRACE("batch size 1");
     testDarknetModel(config_file, weights_file, ref.rowRange(0, 3), scoreDiff, iouDiff);
     }
 
-#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_LE(2018050000)
+#if defined(INF_ENGINE_RELEASE)
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL)
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL)  // Test with 'batch size 2' is disabled for DLIE/OpenCL target
-#endif
-
-#if defined(INF_ENGINE_RELEASE) && INF_ENGINE_VER_MAJOR_EQ(2019020000)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+    {
+        if (INF_ENGINE_VER_MAJOR_LE(2018050000) && target == DNN_TARGET_OPENCL)
+            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL, CV_TEST_TAG_DNN_SKIP_IE_2018R5);
+        else if (INF_ENGINE_VER_MAJOR_EQ(2019020000))
         {
             if (target == DNN_TARGET_OPENCL)
                 applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL, CV_TEST_TAG_DNN_SKIP_IE_2019R2);
             if (target == DNN_TARGET_OPENCL_FP16)
                 applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16, CV_TEST_TAG_DNN_SKIP_IE_2019R2);
         }
+        else if (target == DNN_TARGET_MYRIAD &&
+                 getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+            applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);
+    }
 #endif
 
     {