Enable more deep learning tests

modules/dnn/src/dnn.cpp

@@ -2729,9 +2729,9 @@ void Layer::applyHalideScheduler(Ptr<BackendNode>& node, const std::vector<Mat*>
     }
     else if (targetId == DNN_TARGET_OPENCL)
     {
-        int c_split = outC > 8 ? (outC > 16 ? 8 : 4) : outC;
         if (outW == 1 && outH == 1)
         {
+            int c_split = outC > 8 ? (outC > 16 ? 8 : 4) : outC;
             top.split(c, co, ci, c_split)
                .fuse(x, y, tile).fuse(co, tile, tile).fuse(n, tile, tile)
                .gpu_blocks(tile)
@@ -2741,6 +2741,8 @@ void Layer::applyHalideScheduler(Ptr<BackendNode>& node, const std::vector<Mat*>
         {
             int x_split = outW > 8 ? (outW >= 32 ? 16 : 8) : outW;
             int y_split = outH > 8 ? (outH >= 32 ? 16 : 8) : outH;
+            // Supported vectorization widths: 2, 3, 4, 8, 16
+            int c_split = outC > 8 ? (outC > 16 ? 8 : 4) : std::min(4, outC);
             top.split(x, xo, xi, x_split).split(y, yo, yi, y_split)
                .split(c, co, ci, c_split)
                .gpu_blocks(xo, yo, co)

modules/dnn/src/layers/convolution_layer.cpp

@@ -82,7 +82,21 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
         if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
-            return preferableTarget != DNN_TARGET_MYRIAD || type != "Deconvolution" || adjustPad == Size();
+        {
+            if (type == "Convolution")
+                return preferableTarget != DNN_TARGET_MYRIAD || dilation.width == dilation.height;
+            else
+            {
+                CV_Assert(type == "Deconvolution");
+                const int outGroupCn = blobs[0].size[1];  // Weights are in IOHW layout
+                const int group = numOutput / outGroupCn;
+                if (group != 1)
+                    return false;
+                if (preferableTarget == DNN_TARGET_OPENCL || preferableTarget == DNN_TARGET_OPENCL_FP16)
+                    return dilation.width == 1 && dilation.height == 1;
+                return true;
+            }
+        }
         else
             return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
     }

modules/dnn/src/layers/eltwise_layer.cpp

@@ -97,8 +97,8 @@ public:
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
         return backendId == DNN_BACKEND_OPENCV ||
-               backendId == DNN_BACKEND_HALIDE && haveHalide() ||
-               backendId == DNN_BACKEND_INFERENCE_ENGINE && haveInfEngine();
+               backendId == DNN_BACKEND_HALIDE ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE && (op != SUM || coeffs.empty());
     }
 
     bool getMemoryShapes(const std::vector<MatShape> &inputs,

modules/dnn/src/layers/reorg_layer.cpp

@@ -41,9 +41,9 @@
 //M*/
 
 #include "../precomp.hpp"
+#include "../op_inf_engine.hpp"
 #include <opencv2/dnn/shape_utils.hpp>
 #include <opencv2/dnn/all_layers.hpp>
-#include <iostream>
 
 #ifdef HAVE_OPENCL
 #include "opencl_kernels_dnn.hpp"
@@ -85,6 +85,11 @@ public:
         return false;
     }
 
+    virtual bool supportBackend(int backendId) CV_OVERRIDE
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_INFERENCE_ENGINE;
+    }
+
 #ifdef HAVE_OPENCL
     bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
@@ -169,6 +174,20 @@ public:
         }
     }
 
+    virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
+    {
+#ifdef HAVE_INF_ENGINE
+        InferenceEngine::LayerParams lp;
+        lp.name = name;
+        lp.type = "ReorgYolo";
+        lp.precision = InferenceEngine::Precision::FP32;
+        std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
+        ieLayer->params["stride"] = format("%d", reorgStride);
+        return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
+#endif  // HAVE_INF_ENGINE
+        return Ptr<BackendNode>();
+    }
+
     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
     {

modules/dnn/src/layers/resize_layer.cpp

@@ -192,6 +192,11 @@ public:
         return (outputs[0][2] == inputs[0][2]) && (outputs[0][3] == inputs[0][3]);
     }
 
+    virtual bool supportBackend(int backendId) CV_OVERRIDE
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_INFERENCE_ENGINE;
+    }
+
     virtual void finalize(const std::vector<Mat*>& inputs, std::vector<Mat> &outputs) CV_OVERRIDE
     {
         if (!outWidth && !outHeight)
@@ -204,6 +209,22 @@ public:
         scaleHeight = (outHeight > 1) ? (static_cast<float>(inpHeight - 1) / (outHeight - 1)) : 0.f;
         scaleWidth = (outWidth > 1) ? (static_cast<float>(inpWidth - 1) / (outWidth - 1)) : 0.f;
     }
+
+    virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
+    {
+#ifdef HAVE_INF_ENGINE
+        InferenceEngine::LayerParams lp;
+        lp.name = name;
+        lp.type = "Interp";
+        lp.precision = InferenceEngine::Precision::FP32;
+
+        std::shared_ptr<InferenceEngine::CNNLayer> ieLayer(new InferenceEngine::CNNLayer(lp));
+        ieLayer->params["pad_beg"] = "0";
+        ieLayer->params["pad_end"] = "0";
+        return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
+#endif  // HAVE_INF_ENGINE
+        return Ptr<BackendNode>();
+    }
 };
 
 Ptr<Layer> InterpLayer::create(const LayerParams& params)

modules/dnn/src/layers/slice_layer.cpp

@@ -266,7 +266,21 @@ public:
         std::shared_ptr<InferenceEngine::CropLayer> ieLayer(new InferenceEngine::CropLayer(lp));
 
         CV_Assert(sliceRanges.size() == 1);
-        for (int i = sliceRanges[0].size() - 1; i >= 0; --i)
+
+        int from, to, step;
+        if (preferableTarget == DNN_TARGET_MYRIAD)
+        {
+            from = 1;
+            to = sliceRanges[0].size() + 1;
+            step = 1;
+        }
+        else
+        {
+            from = sliceRanges[0].size() - 1;
+            to = -1;
+            step = -1;
+        }
+        for (int i = from; i != to; i += step)
         {
             ieLayer->axis.push_back(i);
             ieLayer->offset.push_back(sliceRanges[0][i].start);

modules/dnn/test/test_backends.cpp

@@ -10,18 +10,9 @@
 
 namespace opencv_test { namespace {
 
-class DNNTestNetwork : public TestWithParam <tuple<DNNBackend, DNNTarget> >
+class DNNTestNetwork : public DNNTestLayer
 {
 public:
-    dnn::Backend backend;
-    dnn::Target target;
-
-    DNNTestNetwork()
-    {
-        backend = (dnn::Backend)(int)get<0>(GetParam());
-        target = (dnn::Target)(int)get<1>(GetParam());
-    }
-
     void processNet(const std::string& weights, const std::string& proto,
                     Size inpSize, const std::string& outputLayer = "",
                     const std::string& halideScheduler = "",
@@ -40,32 +31,10 @@ public:
                     std::string halideScheduler = "",
                     double l1 = 0.0, double lInf = 0.0, double detectionConfThresh = 0.2)
     {
-        if (backend == DNN_BACKEND_OPENCV && (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
-        {
-#ifdef HAVE_OPENCL
-            if (!cv::ocl::useOpenCL())
-#endif
-            {
-                throw SkipTestException("OpenCL is not available/disabled in OpenCV");
-            }
-        }
-        if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
-        {
-            if (!checkMyriadTarget())
-            {
-                throw SkipTestException("Myriad is not available/disabled in OpenCV");
-            }
-        }
-        if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
-        {
-            l1 = l1 == 0.0 ? 4e-3 : l1;
-            lInf = lInf == 0.0 ? 2e-2 : lInf;
-        }
-        else
-        {
-            l1 = l1 == 0.0 ? 1e-5 : l1;
-            lInf = lInf == 0.0 ? 1e-4 : lInf;
-        }
+        checkBackend();
+        l1 = l1 ? l1 : default_l1;
+        lInf = lInf ? lInf : default_lInf;
+
         weights = findDataFile(weights, false);
         if (!proto.empty())
             proto = findDataFile(proto, false);

modules/dnn/test/test_precomp.hpp

@@ -69,6 +69,93 @@ static testing::internal::ParamGenerator<DNNTarget> availableDnnTargets()
     return testing::ValuesIn(targets);
 }
 
+static testing::internal::ParamGenerator<tuple<DNNBackend, DNNTarget> > dnnBackendsAndTargets()
+{
+    static const tuple<DNNBackend, DNNTarget> testCases[] = {
+    #ifdef HAVE_INF_ENGINE
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_INFERENCE_ENGINE, DNN_TARGET_CPU),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_INFERENCE_ENGINE, DNN_TARGET_OPENCL),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_INFERENCE_ENGINE, DNN_TARGET_OPENCL_FP16),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_INFERENCE_ENGINE, DNN_TARGET_MYRIAD),
+    #endif
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_OPENCV, DNN_TARGET_CPU),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16)
+    };
+    return testing::ValuesIn(testCases);
+}
+
+class DNNTestLayer : public TestWithParam <tuple<DNNBackend, DNNTarget> >
+{
+public:
+    dnn::Backend backend;
+    dnn::Target target;
+    double default_l1, default_lInf;
+
+    DNNTestLayer()
+    {
+        backend = (dnn::Backend)(int)get<0>(GetParam());
+        target = (dnn::Target)(int)get<1>(GetParam());
+        getDefaultThresholds(backend, target, &default_l1, &default_lInf);
+    }
+
+   static void getDefaultThresholds(int backend, int target, double* l1, double* lInf)
+   {
+       if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+       {
+           *l1 = 4e-3;
+           *lInf = 2e-2;
+       }
+       else
+       {
+           *l1 = 1e-5;
+           *lInf = 1e-4;
+       }
+   }
+
+   static void checkBackend(int backend, int target, Mat* inp = 0, Mat* ref = 0)
+   {
+       if (backend == DNN_BACKEND_OPENCV && (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
+       {
+#ifdef HAVE_OPENCL
+           if (!cv::ocl::useOpenCL())
+#endif
+           {
+               throw SkipTestException("OpenCL is not available/disabled in OpenCV");
+           }
+       }
+       if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+       {
+           if (!checkMyriadTarget())
+           {
+               throw SkipTestException("Myriad is not available/disabled in OpenCV");
+           }
+           if (inp && ref && inp->size[0] != 1)
+           {
+               // Myriad plugin supports only batch size 1. Slice a single sample.
+               if (inp->size[0] == ref->size[0])
+               {
+                   std::vector<cv::Range> range(inp->dims, Range::all());
+                   range[0] = Range(0, 1);
+                   *inp = inp->operator()(range);
+
+                   range = std::vector<cv::Range>(ref->dims, Range::all());
+                   range[0] = Range(0, 1);
+                   *ref = ref->operator()(range);
+               }
+               else
+                   throw SkipTestException("Myriad plugin supports only batch size 1");
+           }
+       }
+   }
+
+protected:
+    void checkBackend(Mat* inp = 0, Mat* ref = 0)
+    {
+        checkBackend(backend, target, inp, ref);
+    }
+};
+
 }}
 
 #endif

modules/dnn/test/test_torch_importer.cpp

@@ -296,7 +296,6 @@ TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
         Mat inputBlob = blobFromImage(img, 1.0, Size(), Scalar(103.939, 116.779, 123.68), false);
 
         net.setInput(inputBlob);
-        net.setPreferableBackend(DNN_BACKEND_OPENCV);
         Mat out = net.forward();
 
         // Deprocessing.