Enable more deep learning tests
authorDmitry Kurtaev <dmitry.kurtaev+github@gmail.com>
Wed, 27 Jun 2018 13:34:36 +0000 (16:34 +0300)
committerDmitry Kurtaev <dmitry.kurtaev+github@gmail.com>
Thu, 5 Jul 2018 11:23:15 +0000 (14:23 +0300)
13 files changed:
modules/dnn/src/dnn.cpp
modules/dnn/src/layers/convolution_layer.cpp
modules/dnn/src/layers/eltwise_layer.cpp
modules/dnn/src/layers/reorg_layer.cpp
modules/dnn/src/layers/resize_layer.cpp
modules/dnn/src/layers/slice_layer.cpp
modules/dnn/test/test_backends.cpp
modules/dnn/test/test_darknet_importer.cpp
modules/dnn/test/test_halide_layers.cpp
modules/dnn/test/test_layers.cpp
modules/dnn/test/test_precomp.hpp
modules/dnn/test/test_tf_importer.cpp
modules/dnn/test/test_torch_importer.cpp

index 0177b31..469951c 100644 (file)
@@ -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)
index 27818e5..61489e7 100644 (file)
@@ -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;
     }
index 07a7f60..8eb3ff4 100644 (file)
@@ -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,
index 102f298..89b6f1d 100644 (file)
@@ -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
     {
index 358ee8d..4bf7b50 100644 (file)
@@ -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)
index f6f4109..e4c723e 100644 (file)
@@ -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);
index 48fe765..ad7eb09 100644 (file)
 
 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);
index 2232aa4..682213b 100644 (file)
@@ -65,76 +65,84 @@ TEST(Test_Darknet, read_yolo_voc)
     ASSERT_FALSE(net.empty());
 }
 
-// Test object detection network from Darknet framework.
-static void testDarknetModel(const std::string& cfg, const std::string& weights,
-                             const std::vector<cv::String>& outNames,
-                             const std::vector<int>& refClassIds,
-                             const std::vector<float>& refConfidences,
-                             const std::vector<Rect2d>& refBoxes,
-                             int backendId, int targetId, float scoreDiff = 0.0,
-                             float iouDiff = 0.0, float confThreshold = 0.24)
+class Test_Darknet_layers : public DNNTestLayer
 {
-    if (backendId == DNN_BACKEND_OPENCV && targetId == DNN_TARGET_OPENCL)
+public:
+    void testDarknetLayer(const std::string& name, bool hasWeights = false)
     {
-  #ifdef HAVE_OPENCL
-        if (!cv::ocl::useOpenCL())
-  #endif
-        {
-            throw SkipTestException("OpenCL is not available/disabled in OpenCV");
-        }
-    }
-    if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD)
-    {
-        if (!checkMyriadTarget())
-        {
-            throw SkipTestException("Myriad is not available/disabled in OpenCV");
-        }
+        std::string cfg = findDataFile("dnn/darknet/" + name + ".cfg", false);
+        std::string model = "";
+        if (hasWeights)
+            model = findDataFile("dnn/darknet/" + name + ".weights", false);
+        Mat inp = blobFromNPY(findDataFile("dnn/darknet/" + name + "_in.npy", false));
+        Mat ref = blobFromNPY(findDataFile("dnn/darknet/" + name + "_out.npy", false));
+
+        checkBackend(&inp, &ref);
+
+        Net net = readNet(cfg, model);
+        net.setPreferableBackend(backend);
+        net.setPreferableTarget(target);
+        net.setInput(inp);
+        Mat out = net.forward();
+        normAssert(out, ref, "", default_l1, default_lInf);
     }
-    Mat sample = imread(_tf("dog416.png"));
-    Mat inp = blobFromImage(sample, 1.0/255, Size(416, 416), Scalar(), true, false);
-
-    Net net = readNet(findDataFile("dnn/" + cfg, false),
-                      findDataFile("dnn/" + weights, false));
-    net.setPreferableBackend(backendId);
-    net.setPreferableTarget(targetId);
-    net.setInput(inp);
-    std::vector<Mat> outs;
-    net.forward(outs, outNames);
-
-    std::vector<int> classIds;
-    std::vector<float> confidences;
-    std::vector<Rect2d> boxes;
-    for (int i = 0; i < outs.size(); ++i)
+};
+
+class Test_Darknet_nets : public DNNTestLayer
+{
+public:
+    // Test object detection network from Darknet framework.
+    void testDarknetModel(const std::string& cfg, const std::string& weights,
+                          const std::vector<cv::String>& outNames,
+                          const std::vector<int>& refClassIds,
+                          const std::vector<float>& refConfidences,
+                          const std::vector<Rect2d>& refBoxes,
+                          double scoreDiff, double iouDiff, float confThreshold = 0.24)
     {
-        Mat& out = outs[i];
-        for (int j = 0; j < out.rows; ++j)
+        checkBackend();
+
+        Mat sample = imread(_tf("dog416.png"));
+        Mat inp = blobFromImage(sample, 1.0/255, Size(416, 416), Scalar(), true, false);
+
+        Net net = readNet(findDataFile("dnn/" + cfg, false),
+                          findDataFile("dnn/" + weights, false));
+        net.setPreferableBackend(backend);
+        net.setPreferableTarget(target);
+        net.setInput(inp);
+        std::vector<Mat> outs;
+        net.forward(outs, outNames);
+
+        std::vector<int> classIds;
+        std::vector<float> confidences;
+        std::vector<Rect2d> boxes;
+        for (int i = 0; i < outs.size(); ++i)
         {
-            Mat scores = out.row(j).colRange(5, out.cols);
-            double confidence;
-            Point maxLoc;
-            minMaxLoc(scores, 0, &confidence, 0, &maxLoc);
-
-            float* detection = out.ptr<float>(j);
-            double centerX = detection[0];
-            double centerY = detection[1];
-            double width = detection[2];
-            double height = detection[3];
-            boxes.push_back(Rect2d(centerX - 0.5 * width, centerY - 0.5 * height,
-                                   width, height));
-            confidences.push_back(confidence);
-            classIds.push_back(maxLoc.x);
+            Mat& out = outs[i];
+            for (int j = 0; j < out.rows; ++j)
+            {
+                Mat scores = out.row(j).colRange(5, out.cols);
+                double confidence;
+                Point maxLoc;
+                minMaxLoc(scores, 0, &confidence, 0, &maxLoc);
+
+                float* detection = out.ptr<float>(j);
+                double centerX = detection[0];
+                double centerY = detection[1];
+                double width = detection[2];
+                double height = detection[3];
+                boxes.push_back(Rect2d(centerX - 0.5 * width, centerY - 0.5 * height,
+                                       width, height));
+                confidences.push_back(confidence);
+                classIds.push_back(maxLoc.x);
+            }
         }
+        normAssertDetections(refClassIds, refConfidences, refBoxes, classIds,
+                             confidences, boxes, "", confThreshold, scoreDiff, iouDiff);
     }
-    normAssertDetections(refClassIds, refConfidences, refBoxes, classIds,
-                         confidences, boxes, "", confThreshold, scoreDiff, iouDiff);
-}
-
-typedef testing::TestWithParam<tuple<DNNBackend, DNNTarget> > Test_Darknet_nets;
+};
 
 TEST_P(Test_Darknet_nets, YoloVoc)
 {
-    int backendId = get<0>(GetParam());
-    int targetId = get<1>(GetParam());
     std::vector<cv::String> outNames(1, "detection_out");
 
     std::vector<int> classIds(3);
@@ -143,34 +151,28 @@ TEST_P(Test_Darknet_nets, YoloVoc)
     classIds[0] = 6;  confidences[0] = 0.750469f; boxes[0] = Rect2d(0.577374, 0.127391, 0.325575, 0.173418);  // a car
     classIds[1] = 1;  confidences[1] = 0.780879f; boxes[1] = Rect2d(0.270762, 0.264102, 0.461713, 0.48131); // a bicycle
     classIds[2] = 11; confidences[2] = 0.901615f; boxes[2] = Rect2d(0.1386, 0.338509, 0.282737, 0.60028);  // a dog
-    double scoreDiff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 1e-2 : 8e-5;
-    double iouDiff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.013 : 3e-5;
+    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1e-2 : 8e-5;
+    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.013 : 3e-5;
     testDarknetModel("yolo-voc.cfg", "yolo-voc.weights", outNames,
-                     classIds, confidences, boxes, backendId, targetId, scoreDiff, iouDiff);
+                     classIds, confidences, boxes, scoreDiff, iouDiff);
 }
 
 TEST_P(Test_Darknet_nets, TinyYoloVoc)
 {
-    int backendId = get<0>(GetParam());
-    int targetId = get<1>(GetParam());
     std::vector<cv::String> outNames(1, "detection_out");
     std::vector<int> classIds(2);
     std::vector<float> confidences(2);
     std::vector<Rect2d> boxes(2);
     classIds[0] = 6;  confidences[0] = 0.761967f; boxes[0] = Rect2d(0.579042, 0.159161, 0.31544, 0.160779);  // a car
     classIds[1] = 11; confidences[1] = 0.780595f; boxes[1] = Rect2d(0.129696, 0.386467, 0.315579, 0.534527);  // a dog
-    double scoreDiff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 8e-3 : 8e-5;
-    double iouDiff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 8e-3 : 3e-5;
+    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 8e-3 : 8e-5;
+    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 8e-3 : 3e-5;
     testDarknetModel("tiny-yolo-voc.cfg", "tiny-yolo-voc.weights", outNames,
-                     classIds, confidences, boxes, backendId, targetId, scoreDiff, iouDiff);
+                     classIds, confidences, boxes, scoreDiff, iouDiff);
 }
 
 TEST_P(Test_Darknet_nets, YOLOv3)
 {
-    int backendId = get<0>(GetParam());
-    int targetId = get<1>(GetParam());
-    if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD)
-        throw SkipTestException("");
     std::vector<cv::String> outNames(3);
     outNames[0] = "yolo_82";
     outNames[1] = "yolo_94";
@@ -182,55 +184,41 @@ TEST_P(Test_Darknet_nets, YOLOv3)
     classIds[0] = 7;  confidences[0] = 0.952983f; boxes[0] = Rect2d(0.614622, 0.150257, 0.286747, 0.138994);  // a truck
     classIds[1] = 1; confidences[1] = 0.987908f; boxes[1] = Rect2d(0.150913, 0.221933, 0.591342, 0.524327);  // a bicycle
     classIds[2] = 16; confidences[2] = 0.998836f; boxes[2] = Rect2d(0.160024, 0.389964, 0.257861, 0.553752);  // a dog (COCO)
-    double scoreDiff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 4e-3 : 8e-5;
-    double iouDiff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.011 : 3e-5;
+    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 4e-3 : 8e-5;
+    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.011 : 3e-5;
     testDarknetModel("yolov3.cfg", "yolov3.weights", outNames,
-                     classIds, confidences, boxes, backendId, targetId, scoreDiff, iouDiff);
+                     classIds, confidences, boxes, scoreDiff, iouDiff);
 }
 
-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)
-};
+INSTANTIATE_TEST_CASE_P(/**/, Test_Darknet_nets, dnnBackendsAndTargets());
 
-INSTANTIATE_TEST_CASE_P(/**/, Test_Darknet_nets, testing::ValuesIn(testCases));
+TEST_P(Test_Darknet_layers, shortcut)
+{
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_CPU)
+        throw SkipTestException("");
+    testDarknetLayer("shortcut");
+}
 
-static void testDarknetLayer(const std::string& name, bool hasWeights = false)
+TEST_P(Test_Darknet_layers, upsample)
 {
-    std::string cfg = findDataFile("dnn/darknet/" + name + ".cfg", false);
-    std::string model = "";
-    if (hasWeights)
-        model = findDataFile("dnn/darknet/" + name + ".weights", false);
-    Mat inp = blobFromNPY(findDataFile("dnn/darknet/" + name + "_in.npy", false));
-    Mat ref = blobFromNPY(findDataFile("dnn/darknet/" + name + "_out.npy", false));
-
-    Net net = readNet(cfg, model);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    net.setInput(inp);
-    Mat out = net.forward();
-    normAssert(out, ref);
+    testDarknetLayer("upsample");
 }
 
-TEST(Test_Darknet, shortcut)
+TEST_P(Test_Darknet_layers, avgpool_softmax)
 {
-    testDarknetLayer("shortcut");
+    testDarknetLayer("avgpool_softmax");
 }
 
-TEST(Test_Darknet, upsample)
+TEST_P(Test_Darknet_layers, region)
 {
-    testDarknetLayer("upsample");
+    testDarknetLayer("region");
 }
 
-TEST(Test_Darknet, avgpool_softmax)
+TEST_P(Test_Darknet_layers, reorg)
 {
-    testDarknetLayer("avgpool_softmax");
+    testDarknetLayer("reorg");
 }
 
+INSTANTIATE_TEST_CASE_P(/**/, Test_Darknet_layers, dnnBackendsAndTargets());
+
 }} // namespace
index 2d137c5..b21ae85 100644 (file)
 
 namespace opencv_test { namespace {
 
-#ifdef HAVE_HALIDE
 using namespace cv;
 using namespace cv::dnn;
 using namespace testing;
 
-static void test(LayerParams& params, Mat& input)
+static void test(Mat& input, Net& net, int backendId, int targetId)
 {
+    DNNTestLayer::checkBackend(backendId, targetId);
     randu(input, -1.0f, 1.0f);
 
-    Net net;
-    int lid = net.addLayer(params.name, params.type, params);
-    net.connect(0, 0, lid, 0);
-
     net.setInput(input);
     net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    Mat outputDefault = net.forward(params.name).clone();
+    Mat outputDefault = net.forward().clone();
 
-    net.setPreferableBackend(DNN_BACKEND_HALIDE);
-    Mat outputHalide = net.forward(params.name).clone();
-    normAssert(outputDefault, outputHalide);
+    net.setPreferableBackend(backendId);
+    net.setPreferableTarget(targetId);
+    Mat outputHalide = net.forward().clone();
+
+    double l1, lInf;
+    DNNTestLayer::getDefaultThresholds(backendId, targetId, &l1, &lInf);
+    normAssert(outputDefault, outputHalide, "", l1, lInf);
+}
+
+static void test(LayerParams& params, Mat& input, int backendId, int targetId)
+{
+    Net net;
+    net.addLayerToPrev(params.name, params.type, params);
+    test(input, net, backendId, targetId);
+}
+
+static testing::internal::ParamGenerator<tuple<DNNBackend, DNNTarget> > dnnBackendsAndTargetsWithHalide()
+{
+    static const tuple<DNNBackend, DNNTarget> testCases[] = {
+#ifdef HAVE_HALIDE
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_HALIDE, DNN_TARGET_CPU),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_HALIDE, DNN_TARGET_OPENCL),
+#endif
+#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_OPENCL),
+        tuple<DNNBackend, DNNTarget>(DNN_BACKEND_OPENCV, DNN_TARGET_OPENCL_FP16)
+    };
+    return testing::ValuesIn(testCases);
 }
 
+class Test_Halide_layers : public DNNTestLayer {};
+
 ////////////////////////////////////////////////////////////////////////////////
 // Padding
 ////////////////////////////////////////////////////////////////////////////////
-TEST(Padding_Halide, Accuracy)
+TEST_P(Test_Halide_layers, Padding)
 {
     static const int kNumRuns = 10;
     std::vector<int> paddings(8);
@@ -52,15 +80,16 @@ TEST(Padding_Halide, Accuracy)
         lp.type = "Padding";
         lp.name = "testLayer";
 
-        Mat input({1 + rng(10), 1 + rng(10), 1 + rng(10), 1 + rng(10)}, CV_32F);
-        test(lp, input);
+        int sz[] = {1 + (int)rng(10), 1 + (int)rng(10), 1 + (int)rng(10), 1 + (int)rng(10)};
+        Mat input(4, &sz[0], CV_32F);
+        test(lp, input, backend, target);
     }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 // Convolution
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<Vec3i, Size, Size, Size, Size, Size, bool> > Convolution;
+typedef TestWithParam<tuple<Vec3i, Size, Size, Size, Size, Size, bool, tuple<DNNBackend, DNNTarget> > > Convolution;
 TEST_P(Convolution, Accuracy)
 {
     int inChannels = get<0>(GetParam())[0];
@@ -72,8 +101,15 @@ TEST_P(Convolution, Accuracy)
     Size pad = get<4>(GetParam());
     Size dilation = get<5>(GetParam());
     bool hasBias = get<6>(GetParam());
+    int backendId = get<0>(get<7>(GetParam()));
+    int targetId = get<1>(get<7>(GetParam()));
+
+    if ((backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD) ||
+        (backendId == DNN_BACKEND_OPENCV && targetId == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
 
-    Mat weights({outChannels, inChannels / group, kernel.height, kernel.width}, CV_32F);
+    int sz[] = {outChannels, inChannels / group, kernel.height, kernel.width};
+    Mat weights(4, &sz[0], CV_32F);
     randu(weights, -1.0f, 1.0f);
 
     LayerParams lp;
@@ -93,12 +129,13 @@ TEST_P(Convolution, Accuracy)
     lp.blobs.push_back(weights);
     if (hasBias)
     {
-        Mat bias({outChannels}, CV_32F);
+        Mat bias(1, outChannels, CV_32F);
         randu(bias, -1.0f, 1.0f);
         lp.blobs.push_back(bias);
     }
-    Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
-    test(lp, input);
+    int inpSz[] = {1, inChannels, inSize.height, inSize.width};
+    Mat input(4, &inpSz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Convolution, Combine(
@@ -110,13 +147,14 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Convolution, Combine(
 /*stride*/   Values(Size(1, 1), Size(2, 2)),
 /*pad*/      Values(Size(1, 0), Size(0, 1)),
 /*dilation*/ Values(Size(1, 1), Size(2, 2)),
-/*has bias*/ Bool()
+/*has bias*/ Bool(),
+             dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
 // Deconvolution
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<Vec3i, Size, Size, Size, Size, Vec4i, bool> > Deconvolution;
+typedef TestWithParam<tuple<Vec3i, Size, Size, Size, Size, Vec4i, bool, tuple<DNNBackend, DNNTarget> > > Deconvolution;
 TEST_P(Deconvolution, Accuracy)
 {
     int inChannels = get<0>(GetParam())[0];
@@ -129,8 +167,14 @@ TEST_P(Deconvolution, Accuracy)
     Size stride = Size(get<5>(GetParam())[0], get<5>(GetParam())[1]);
     Size adjPad = Size(get<5>(GetParam())[2], get<5>(GetParam())[3]);
     bool hasBias = get<6>(GetParam());
-
-    Mat weights({inChannels, outChannels / group, kernel.height, kernel.width}, CV_32F);
+    int backendId = get<0>(get<7>(GetParam()));
+    int targetId = get<1>(get<7>(GetParam()));
+    if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_CPU &&
+        dilation.width == 2 && dilation.height == 2)
+        throw SkipTestException("");
+
+    int sz[] = {inChannels, outChannels / group, kernel.height, kernel.width};
+    Mat weights(4, &sz[0], CV_32F);
     randu(weights, -1.0f, 1.0f);
 
     LayerParams lp;
@@ -152,12 +196,13 @@ TEST_P(Deconvolution, Accuracy)
     lp.blobs.push_back(weights);
     if (hasBias)
     {
-        Mat bias({outChannels}, CV_32F);
+        Mat bias(1, outChannels, CV_32F);
         randu(bias, -1.0f, 1.0f);
         lp.blobs.push_back(bias);
     }
-    Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
-    test(lp, input);
+    int inpSz[] = {1, inChannels, inSize.height, inSize.width};
+    Mat input(4, &inpSz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Deconvolution, Combine(
@@ -168,13 +213,14 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Deconvolution, Combine(
 /*pad*/      Values(Size(1, 0), Size(0, 1)),
 /*dilation*/ Values(Size(1, 1), Size(2, 2)),
 /*stride, adj. pad*/ Values(Vec4i(1,1, 0,0), Vec4i(2,2, 1,0), Vec4i(1,2, 0,1)),
-/*has bias*/ Bool()
+/*has bias*/ Bool(),
+             dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
 // LRN
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<Vec3i, int, Vec3f, bool, std::string> > LRN;
+typedef TestWithParam<tuple<Vec3i, int, Vec3f, bool, std::string, tuple<DNNBackend, DNNTarget> > > LRN;
 TEST_P(LRN, Accuracy)
 {
     int inChannels = get<0>(GetParam())[0];
@@ -185,6 +231,10 @@ TEST_P(LRN, Accuracy)
     float bias = get<2>(GetParam())[2];
     bool normBySize = get<3>(GetParam());
     std::string nrmType = get<4>(GetParam());
+    int backendId = get<0>(get<5>(GetParam()));
+    int targetId = get<1>(get<5>(GetParam()));
+    if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
 
     LayerParams lp;
     lp.set("norm_region", nrmType);
@@ -196,8 +246,9 @@ TEST_P(LRN, Accuracy)
     lp.type = "LRN";
     lp.name = "testLayer";
 
-    Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
-    test(lp, input);
+    int sz[] = {1, inChannels, inSize.height, inSize.width};
+    Mat input(4, &sz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, LRN, Combine(
@@ -207,19 +258,24 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, LRN, Combine(
 /*alpha, beta,*/        Vec3f(1.0f, 0.9f, 1.1f), Vec3f(1.0f, 1.1f, 0.9f),
 /*bias */               Vec3f(1.1f, 0.9f, 1.0f), Vec3f(1.1f, 1.0f, 0.9f)),
 /*norm_by_size*/ Bool(),
-/*norm_type*/    Values("ACROSS_CHANNELS", "WITHIN_CHANNEL")
+/*norm_type*/    Values("ACROSS_CHANNELS", "WITHIN_CHANNEL"),
+                 dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
 // Average pooling
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<int, Size, Size, Size> > AvePooling;
+typedef TestWithParam<tuple<int, Size, Size, Size, tuple<DNNBackend, DNNTarget> > > AvePooling;
 TEST_P(AvePooling, Accuracy)
 {
     int inChannels = get<0>(GetParam());
     Size outSize = get<1>(GetParam());;  // Input size will be computed from parameters.
     Size kernel = get<2>(GetParam());
     Size stride = get<3>(GetParam());
+    int backendId = get<0>(get<4>(GetParam()));
+    int targetId = get<1>(get<4>(GetParam()));
+    if (backendId == DNN_BACKEND_INFERENCE_ENGINE && targetId == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
 
     const int inWidth = (outSize.width - 1) * stride.width + kernel.width;
     const int inHeight = (outSize.height - 1) * stride.height + kernel.height;
@@ -233,21 +289,23 @@ TEST_P(AvePooling, Accuracy)
     lp.type = "Pooling";
     lp.name = "testLayer";
 
-    Mat input({1, inChannels, inHeight, inWidth}, CV_32F);
-    test(lp, input);
+    int sz[] = {1, inChannels, inHeight, inWidth};
+    Mat input(4, &sz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, AvePooling, Combine(
 /*in channels*/ Values(3, 4),
 /*out size*/    Values(Size(1, 1), Size(2, 2), Size(3, 2), Size(4, 7)),
 /*kernel*/      Values(Size(1, 1), Size(2, 2), Size(3, 3), Size(3, 2)),
-/*stride*/      Values(Size(1, 1), Size(2, 2), Size(3, 2))
+/*stride*/      Values(Size(1, 1), Size(2, 2), Size(3, 2)),
+                dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
 // Maximum pooling
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<int, Size, Size, Size, Size> > MaxPooling;
+typedef TestWithParam<tuple<int, Size, Size, Size, Size, tuple<DNNBackend, DNNTarget> > > MaxPooling;
 TEST_P(MaxPooling, Accuracy)
 {
     int inChannels = get<0>(GetParam());
@@ -255,6 +313,8 @@ TEST_P(MaxPooling, Accuracy)
     Size kernel = get<2>(GetParam());
     Size stride = get<3>(GetParam());
     Size pad = get<4>(GetParam());
+    int backendId = get<0>(get<5>(GetParam()));
+    int targetId = get<1>(get<5>(GetParam()));
 
     LayerParams lp;
     lp.set("pool", "max");
@@ -267,8 +327,9 @@ TEST_P(MaxPooling, Accuracy)
     lp.type = "Pooling";
     lp.name = "testLayer";
 
-    Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
-    test(lp, input);
+    int sz[] = {1, inChannels, inSize.height, inSize.width};
+    Mat input(4, &sz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, MaxPooling, Combine(
@@ -276,19 +337,25 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, MaxPooling, Combine(
 /*in size*/     Values(Size(5, 5), Size(7, 6)),
 /*kernel*/      Values(Size(2, 2), Size(3, 3), Size(3, 2)),
 /*stride*/      Values(Size(1, 1), Size(2, 2), Size(3, 2)),
-/*pad*/         Values(Size(0, 0), Size(1, 1), Size(0, 1))
+/*pad*/         Values(Size(0, 0), Size(1, 1), Size(0, 1)),
+                dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
 // Fully-connected
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<int, Size, int, bool> > FullyConnected;
+typedef TestWithParam<tuple<int, Size, int, bool, tuple<DNNBackend, DNNTarget> > > FullyConnected;
 TEST_P(FullyConnected, Accuracy)
 {
     int inChannels = get<0>(GetParam());
     Size inSize = get<1>(GetParam());
     int outChannels = get<2>(GetParam());
     bool hasBias = get<3>(GetParam());
+    int backendId = get<0>(get<4>(GetParam()));
+    int targetId = get<1>(get<4>(GetParam()));
+    if (backendId == DNN_BACKEND_INFERENCE_ENGINE ||
+        (backendId == DNN_BACKEND_OPENCV && targetId == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
 
     Mat weights(outChannels, inChannels * inSize.height * inSize.width, CV_32F);
     randu(weights, -1.0f, 1.0f);
@@ -304,39 +371,50 @@ TEST_P(FullyConnected, Accuracy)
     lp.type = "InnerProduct";
     lp.name = "testLayer";
 
-    Mat input({1, inChannels, inSize.height, inSize.width}, CV_32F);
-    test(lp, input);
+    int sz[] = {1, inChannels, inSize.height, inSize.width};
+    Mat input(4, &sz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, FullyConnected, Combine(
 /*in channels*/  Values(3, 4),
 /*in size*/      Values(Size(5, 4), Size(4, 5), Size(1, 1)),
 /*out channels*/ Values(3, 4),
-/*has bias*/     Bool()
+/*has bias*/     Bool(),
+                 dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
 // SoftMax
 ////////////////////////////////////////////////////////////////////////////////
-typedef TestWithParam<tuple<int> > SoftMax;
+typedef TestWithParam<tuple<int,  tuple<DNNBackend, DNNTarget> > > SoftMax;
 TEST_P(SoftMax, Accuracy)
 {
     int inChannels = get<0>(GetParam());
+    int backendId = get<0>(get<1>(GetParam()));
+    int targetId = get<1>(get<1>(GetParam()));
     LayerParams lp;
     lp.type = "SoftMax";
     lp.name = "testLayer";
 
-    Mat input({1, inChannels, 1, 1}, CV_32F);
-    test(lp, input);
+    int sz[] = {1, inChannels, 1, 1};
+    Mat input(4, &sz[0], CV_32F);
+    test(lp, input, backendId, targetId);
 }
 
-INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, SoftMax, Values(3, 4, 5, 1024));
+INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, SoftMax, Combine(
+    Values(3, 4, 5, 1024),
+    dnnBackendsAndTargetsWithHalide()
+));
 
 //////////////////////////////////////////////////////////////////////////////
 // Max pooling - unpooling
 //////////////////////////////////////////////////////////////////////////////
-TEST(MaxPoolUnpool_Halide, Accuracy)
+TEST_P(Test_Halide_layers, MaxPoolUnpool)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
+
     LayerParams pool;
     pool.set("pool", "max");
     pool.set("kernel_w", 2);
@@ -366,16 +444,9 @@ TEST(MaxPoolUnpool_Halide, Accuracy)
     net.connect(poolId, 0, unpoolId, 0);
     net.connect(poolId, 1, unpoolId, 1);
 
-    Mat input({1, 1, 4, 4}, CV_32F);
-    randu(input, -1.0f, 1.0f);
-    net.setInput(input);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    Mat outputDefault = net.forward("testUnpool").clone();
-
-    net.setPreferableBackend(DNN_BACKEND_HALIDE);
-    net.setInput(input);
-    Mat outputHalide = net.forward("testUnpool").clone();
-    normAssert(outputDefault, outputHalide);
+    int sz[] = {1, 1, 4, 4};
+    Mat input(4, &sz[0], CV_32F);
+    test(input, net, backend, target);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -383,7 +454,7 @@ TEST(MaxPoolUnpool_Halide, Accuracy)
 ////////////////////////////////////////////////////////////////////////////////
 static const int kNumChannels = 3;
 
-void testInPlaceActivation(LayerParams& lp)
+void testInPlaceActivation(LayerParams& lp, int backendId, int targetId)
 {
     EXPECT_FALSE(lp.name.empty());
 
@@ -400,24 +471,19 @@ void testInPlaceActivation(LayerParams& lp)
     net.connect(0, 0, poolId, 0);
     net.addLayerToPrev(lp.name, lp.type, lp);
 
-    Mat input({1, kNumChannels, 10, 10}, CV_32F);
-    randu(input, -1.0f, 1.0f);
-    net.setInput(input);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    Mat outputDefault = net.forward(lp.name).clone();
-
-    net.setInput(input);
-    net.setPreferableBackend(DNN_BACKEND_HALIDE);
-    Mat outputHalide = net.forward(lp.name).clone();
-    normAssert(outputDefault, outputHalide);
+    int sz[] = {1, kNumChannels, 10, 10};
+    Mat input(4, &sz[0], CV_32F);
+    test(input, net, backendId, targetId);
 }
 
-typedef TestWithParam<tuple<bool, bool, float> > BatchNorm;
+typedef TestWithParam<tuple<bool, bool, float, tuple<DNNBackend, DNNTarget> > > BatchNorm;
 TEST_P(BatchNorm, Accuracy)
 {
     bool hasWeights = get<0>(GetParam());
     bool hasBias = get<1>(GetParam());
     float epsilon = get<2>(GetParam());
+    int backendId = get<0>(get<3>(GetParam()));
+    int targetId = get<1>(get<3>(GetParam()));
 
     LayerParams lp;
     lp.set("has_weight", hasWeights);
@@ -428,56 +494,66 @@ TEST_P(BatchNorm, Accuracy)
 
     lp.blobs.reserve(4);
     for (int i = 0; i < 3; ++i)
-        lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
+        lp.blobs.push_back(Mat(1, kNumChannels, CV_32F));
     if (hasBias || hasWeights)
-        lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
+        lp.blobs.push_back(Mat(1, kNumChannels, CV_32F));
 
-    for (Mat& m : lp.blobs)
-        randu(m, 0.0f, 1.0f);
+    for (int i = 0; i < lp.blobs.size(); ++i)
+        randu(lp.blobs[i], 0.0f, 1.0f);
 
-    testInPlaceActivation(lp);
+    testInPlaceActivation(lp, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, BatchNorm, Combine(
 /*has weights*/ Bool(),
 /*has bias*/    Bool(),
-/*epsilon*/     Values(1e-3f, 1e-5f)
+/*epsilon*/     Values(1e-3f, 1e-5f),
+                dnnBackendsAndTargetsWithHalide()
 ));
 
-typedef TestWithParam<tuple<float> > ReLU;
+typedef TestWithParam<tuple<float, tuple<DNNBackend, DNNTarget> > > ReLU;
 TEST_P(ReLU, Accuracy)
 {
     float negativeSlope = get<0>(GetParam());
+    int backendId = get<0>(get<1>(GetParam()));
+    int targetId = get<1>(get<1>(GetParam()));
 
     LayerParams lp;
     lp.set("negative_slope", negativeSlope);
     lp.type = "ReLU";
     lp.name = "testLayer";
-    testInPlaceActivation(lp);
+    testInPlaceActivation(lp, backendId, targetId);
 }
 
-INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, ReLU, Values(
-/*negative slope*/ 2.0f, 0.3f, -0.1f, 0.0f
+INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, ReLU, Combine(
+/*negative slope*/ Values(2.0f, 0.3f, -0.1f, 0.0f),
+                   dnnBackendsAndTargetsWithHalide()
 ));
 
-typedef TestWithParam<tuple<std::string> > NoParamActivation;
+typedef TestWithParam<tuple<std::string, tuple<DNNBackend, DNNTarget> > > NoParamActivation;
 TEST_P(NoParamActivation, Accuracy)
 {
+    int backendId = get<0>(get<1>(GetParam()));
+    int targetId = get<1>(get<1>(GetParam()));
+
     LayerParams lp;
     lp.type = get<0>(GetParam());
     lp.name = "testLayer";
-    testInPlaceActivation(lp);
+    testInPlaceActivation(lp, backendId, targetId);
 }
-INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, NoParamActivation, Values(
-/*type*/ "TanH", "Sigmoid", "AbsVal", "BNLL"
+INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, NoParamActivation, Combine(
+/*type*/ Values("TanH", "Sigmoid", "AbsVal", "BNLL"),
+         dnnBackendsAndTargetsWithHalide()
 ));
 
-typedef TestWithParam<tuple<Vec3f> > Power;
+typedef TestWithParam<tuple<Vec3f, tuple<DNNBackend, DNNTarget> > > Power;
 TEST_P(Power, Accuracy)
 {
     float power = get<0>(GetParam())[0];
     float scale = get<0>(GetParam())[1];
     float shift = get<0>(GetParam())[2];
+    int backendId = get<0>(get<1>(GetParam()));
+    int targetId = get<1>(get<1>(GetParam()));
 
     LayerParams lp;
     lp.set("power", power);
@@ -485,46 +561,52 @@ TEST_P(Power, Accuracy)
     lp.set("shift", shift);
     lp.type = "Power";
     lp.name = "testLayer";
-    testInPlaceActivation(lp);
+    testInPlaceActivation(lp, backendId, targetId);
 }
 
-INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Power,
+INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Power, Combine(
 /*power, scale, shift*/ Values(Vec3f(0.9f, 1.0f, 1.1f), Vec3f(0.9f, 1.1f, 1.0f),
                                Vec3f(1.0f, 0.9f, 1.1f), Vec3f(1.0f, 1.1f, 0.9f),
-                               Vec3f(1.1f, 0.9f, 1.0f), Vec3f(1.1f, 1.0f, 0.9f))
-);
+                               Vec3f(1.1f, 0.9f, 1.0f), Vec3f(1.1f, 1.0f, 0.9f)),
+                        dnnBackendsAndTargetsWithHalide()
+));
 
-TEST(ChannelsPReLU, Accuracy)
+TEST_P(Test_Halide_layers, ChannelsPReLU)
 {
     LayerParams lp;
     lp.type = "ChannelsPReLU";
     lp.name = "testLayer";
-    lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
+    lp.blobs.push_back(Mat(1, kNumChannels, CV_32F));
     randu(lp.blobs[0], -1.0f, 1.0f);
 
-    testInPlaceActivation(lp);
+    testInPlaceActivation(lp, backend, target);
 }
 
-typedef TestWithParam<tuple<bool> > Scale;
+typedef TestWithParam<tuple<bool, tuple<DNNBackend, DNNTarget> > > Scale;
 TEST_P(Scale, Accuracy)
 {
     bool hasBias = get<0>(GetParam());
+    int backendId = get<0>(get<1>(GetParam()));
+    int targetId = get<1>(get<1>(GetParam()));
 
     LayerParams lp;
     lp.set("bias_term", hasBias);
     lp.type = "Scale";
     lp.name = "testLayer";
-    lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
+    lp.blobs.push_back(Mat(1, kNumChannels, CV_32F));
     randu(lp.blobs[0], -1.0f, 1.0f);
     if (hasBias)
     {
-        lp.blobs.push_back(Mat({kNumChannels}, CV_32F));
+        lp.blobs.push_back(Mat(1, kNumChannels, CV_32F));
         randu(lp.blobs[1], -1.0f, 1.0f);
     }
-    testInPlaceActivation(lp);
+    testInPlaceActivation(lp, backendId, targetId);
 }
 
-INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Scale, Values(true, false));
+INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Scale, Combine(
+    Bool(),
+    dnnBackendsAndTargetsWithHalide()
+));
 
 ////////////////////////////////////////////////////////////////////////////////
 // Concat layer
@@ -534,11 +616,13 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Scale, Values(true, false));
 //      `--- conv ----^ ^ ^
 //      `---- ... ------' '
 //      `-----------------'
-typedef TestWithParam<tuple<Vec3i, Vec3i> > Concat;
+typedef TestWithParam<tuple<Vec3i, Vec3i, tuple<DNNBackend, DNNTarget> > > Concat;
 TEST_P(Concat, Accuracy)
 {
     Vec3i inSize = get<0>(GetParam());
     Vec3i numChannels = get<1>(GetParam());
+    int backendId = get<0>(get<2>(GetParam()));
+    int targetId = get<1>(get<2>(GetParam()));
 
     Net net;
 
@@ -549,7 +633,8 @@ TEST_P(Concat, Accuracy)
         if (!numChannels[i])
             break;
 
-        Mat weights({numChannels[i], inSize[0], 1, 1}, CV_32F);
+        int sz[] = {numChannels[i], inSize[0], 1, 1};
+        Mat weights(4, &sz[0], CV_32F);
         randu(weights, -1.0f, 1.0f);
 
         LayerParams convParam;
@@ -578,21 +663,15 @@ TEST_P(Concat, Accuracy)
         net.connect(convLayerIds[i], 0, concatId, i + 1);
     }
 
-    Mat input({1, inSize[0], inSize[1], inSize[2]}, CV_32F);
-    randu(input, -1.0f, 1.0f);
-
-    net.setInput(input);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    Mat outputDefault = net.forward(concatParam.name).clone();
-
-    net.setPreferableBackend(DNN_BACKEND_HALIDE);
-    Mat outputHalide = net.forward(concatParam.name).clone();
-    normAssert(outputDefault, outputHalide);
+    int sz[] = {1, inSize[0], inSize[1], inSize[2]};
+    Mat input(4, &sz[0], CV_32F);
+    test(input, net, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Concat, Combine(
 /*input size*/ Values(Vec3i(1, 4, 5), Vec3i(2, 8, 6)),
-/*channels*/   Values(Vec3i(2, 0, 0), Vec3i(3, 4, 0), Vec3i(1, 6, 2))
+/*channels*/   Values(Vec3i(2, 0, 0), Vec3i(3, 4, 0), Vec3i(1, 6, 2)),
+               dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -603,20 +682,27 @@ INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Concat, Combine(
 //      `--- conv ----^ ^ ^
 //      `---- ... ------' '
 //      `-----------------'
-typedef TestWithParam<tuple<Vec3i, std::string, int, bool> > Eltwise;
+typedef TestWithParam<tuple<Vec3i, std::string, int, bool, tuple<DNNBackend, DNNTarget> > > Eltwise;
 TEST_P(Eltwise, Accuracy)
 {
     Vec3i inSize = get<0>(GetParam());
     std::string op = get<1>(GetParam());
     int numConv = get<2>(GetParam());
     bool weighted = get<3>(GetParam());
+    int backendId = get<0>(get<4>(GetParam()));
+    int targetId = get<1>(get<4>(GetParam()));
+
+    if (backendId == DNN_BACKEND_OPENCV &&
+        (targetId == DNN_TARGET_OPENCL || targetId == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
 
     Net net;
 
     std::vector<int> convLayerIds(numConv);
     for (int i = 0; i < numConv; ++i)
     {
-        Mat weights({inSize[0], inSize[0], 1, 1}, CV_32F);
+        int sz[] = {inSize[0], inSize[0], 1, 1};
+        Mat weights(4, &sz[0], CV_32F);
         randu(weights, -1.0f, 1.0f);
 
         LayerParams convParam;
@@ -655,28 +741,23 @@ TEST_P(Eltwise, Accuracy)
         net.connect(convLayerIds[i], 0, eltwiseId, i + 1);
     }
 
-    Mat input({1, inSize[0], inSize[1], inSize[2]}, CV_32F);
-    randu(input, -1.0f, 1.0f);
-
-    net.setInput(input);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    Mat outputDefault = net.forward(eltwiseParam.name).clone();
-
-    net.setPreferableBackend(DNN_BACKEND_HALIDE);
-    Mat outputHalide = net.forward(eltwiseParam.name).clone();
-    normAssert(outputDefault, outputHalide);
+    int sz[] = {1, inSize[0], inSize[1], inSize[2]};
+    Mat input(4, &sz[0], CV_32F);
+    test(input, net, backendId, targetId);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, Eltwise, Combine(
 /*input size*/ Values(Vec3i(1, 4, 5), Vec3i(2, 8, 6)),
 /*operation*/  Values("prod", "sum", "max"),
 /*num convs*/  Values(1, 2, 3),
-/*weighted(for sum only)*/ Bool()
+/*weighted(for sum only)*/ Bool(),
+               dnnBackendsAndTargetsWithHalide()
 ));
 
 ////////////////////////////////////////////////////////////////////////////
 // Mixed backends
 ////////////////////////////////////////////////////////////////////////////
+#ifdef HAVE_HALIDE
 TEST(MixedBackends_Halide_Default_Halide, Accuracy)
 {
     // Just a layer that supports Halide backend.
@@ -700,7 +781,8 @@ TEST(MixedBackends_Halide_Default_Halide, Accuracy)
     net.addLayerToPrev(mvn.name, mvn.type, mvn);
     net.addLayerToPrev(lrn2.name, lrn2.type, lrn2);
 
-    Mat input({4, 3, 5, 6}, CV_32F);
+    int sz[] = {4, 3, 5, 6};
+    Mat input(4, &sz[0], CV_32F);
     randu(input, -1.0f, 1.0f);
     net.setInput(input);
     net.setPreferableBackend(DNN_BACKEND_OPENCV);
@@ -718,4 +800,6 @@ TEST(MixedBackends_Halide_Default_Halide, Accuracy)
 }
 #endif  // HAVE_HALIDE
 
+INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_Halide_layers, dnnBackendsAndTargetsWithHalide());
+
 }} // namespace
index 963206b..798ab84 100644 (file)
@@ -92,75 +92,84 @@ void runLayer(Ptr<Layer> layer, std::vector<Mat> &inpBlobs, std::vector<Mat> &ou
         outBlobs[i] = outp[i];
 }
 
-
-void testLayerUsingCaffeModels(String basename, int targetId = DNN_TARGET_CPU,
-                               bool useCaffeModel = false, bool useCommonInputBlob = true)
+class Test_Caffe_layers : public DNNTestLayer
 {
-    String prototxt = _tf(basename + ".prototxt");
-    String caffemodel = _tf(basename + ".caffemodel");
+public:
+    void testLayerUsingCaffeModels(const String& basename, bool useCaffeModel = false,
+                                   bool useCommonInputBlob = true, double l1 = 0.0,
+                                   double lInf = 0.0)
+    {
+        String prototxt = _tf(basename + ".prototxt");
+        String caffemodel = _tf(basename + ".caffemodel");
 
-    String inpfile = (useCommonInputBlob) ? _tf("blob.npy") : _tf(basename + ".input.npy");
-    String outfile = _tf(basename + ".npy");
+        String inpfile = (useCommonInputBlob) ? _tf("blob.npy") : _tf(basename + ".input.npy");
+        String outfile = _tf(basename + ".npy");
 
-    Net net = readNetFromCaffe(prototxt, (useCaffeModel) ? caffemodel : String());
-    ASSERT_FALSE(net.empty());
+        Mat inp = blobFromNPY(inpfile);
+        Mat ref = blobFromNPY(outfile);
+        checkBackend(&inp, &ref);
 
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    net.setPreferableTarget(targetId);
+        Net net = readNetFromCaffe(prototxt, (useCaffeModel) ? caffemodel : String());
+        ASSERT_FALSE(net.empty());
 
-    Mat inp = blobFromNPY(inpfile);
-    Mat ref = blobFromNPY(outfile);
+        net.setPreferableBackend(backend);
+        net.setPreferableTarget(target);
 
-    net.setInput(inp, "input");
-    Mat out = net.forward("output");
+        net.setInput(inp, "input");
+        Mat out = net.forward("output");
 
-    normAssert(ref, out);
-}
+        normAssert(ref, out, "", l1 ? l1 : default_l1, lInf ? lInf : default_lInf);
+    }
+};
 
-typedef testing::TestWithParam<DNNTarget> Test_Caffe_layers;
 TEST_P(Test_Caffe_layers, Softmax)
 {
-    testLayerUsingCaffeModels("layer_softmax", GetParam());
+    testLayerUsingCaffeModels("layer_softmax");
 }
 
 TEST_P(Test_Caffe_layers, LRN_spatial)
 {
-    testLayerUsingCaffeModels("layer_lrn_spatial", GetParam());
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
+    testLayerUsingCaffeModels("layer_lrn_spatial");
 }
 
 TEST_P(Test_Caffe_layers, LRN_channels)
 {
-    testLayerUsingCaffeModels("layer_lrn_channels", GetParam());
+    testLayerUsingCaffeModels("layer_lrn_channels");
 }
 
 TEST_P(Test_Caffe_layers, Convolution)
 {
-    testLayerUsingCaffeModels("layer_convolution", GetParam(), true);
+    testLayerUsingCaffeModels("layer_convolution", true);
 }
 
 TEST_P(Test_Caffe_layers, DeConvolution)
 {
-    testLayerUsingCaffeModels("layer_deconvolution", GetParam(), true, false);
+    testLayerUsingCaffeModels("layer_deconvolution", true, false);
 }
 
 TEST_P(Test_Caffe_layers, InnerProduct)
 {
-    testLayerUsingCaffeModels("layer_inner_product", GetParam(), true);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
+        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
+    testLayerUsingCaffeModels("layer_inner_product", true);
 }
 
 TEST_P(Test_Caffe_layers, Pooling_max)
 {
-    testLayerUsingCaffeModels("layer_pooling_max", GetParam());
+    testLayerUsingCaffeModels("layer_pooling_max");
 }
 
 TEST_P(Test_Caffe_layers, Pooling_ave)
 {
-    testLayerUsingCaffeModels("layer_pooling_ave", GetParam());
+    testLayerUsingCaffeModels("layer_pooling_ave");
 }
 
 TEST_P(Test_Caffe_layers, MVN)
 {
-    testLayerUsingCaffeModels("layer_mvn", GetParam());
+    testLayerUsingCaffeModels("layer_mvn");
 }
 
 void testReshape(const MatShape& inputShape, const MatShape& targetShape,
@@ -210,33 +219,38 @@ TEST(Layer_Test_Reshape, Accuracy)
     }
 }
 
-TEST(Layer_Test_BatchNorm, Accuracy)
-{
-    testLayerUsingCaffeModels("layer_batch_norm", DNN_TARGET_CPU, true);
-}
-
-TEST(Layer_Test_BatchNorm, local_stats)
+TEST_P(Test_Caffe_layers, BatchNorm)
 {
-    testLayerUsingCaffeModels("layer_batch_norm_local_stats", DNN_TARGET_CPU, true, false);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
+    testLayerUsingCaffeModels("layer_batch_norm", true);
+    testLayerUsingCaffeModels("layer_batch_norm_local_stats", true, false);
 }
 
 TEST_P(Test_Caffe_layers, ReLU)
 {
-    testLayerUsingCaffeModels("layer_relu", GetParam());
+    testLayerUsingCaffeModels("layer_relu");
 }
 
-TEST(Layer_Test_Dropout, Accuracy)
+TEST_P(Test_Caffe_layers, Dropout)
 {
     testLayerUsingCaffeModels("layer_dropout");
 }
 
 TEST_P(Test_Caffe_layers, Concat)
 {
-    testLayerUsingCaffeModels("layer_concat", GetParam());
+    testLayerUsingCaffeModels("layer_concat");
+    testLayerUsingCaffeModels("layer_concat_optim", true, false);
+    testLayerUsingCaffeModels("layer_concat_shared_input", true, false);
 }
 
-TEST(Layer_Test_Fused_Concat, Accuracy)
+TEST_P(Test_Caffe_layers, Fused_Concat)
 {
+    if ((backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_CPU) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL))
+        throw SkipTestException("");
+    checkBackend();
+
     // Test case
     // input
     //   |
@@ -267,28 +281,32 @@ TEST(Layer_Test_Fused_Concat, Accuracy)
     randu(input, 0.0f, 1.0f);  // [0, 1] to make AbsVal an identity transformation.
 
     net.setInput(input);
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
     Mat out = net.forward();
 
-    normAssert(slice(out, Range::all(), Range(0, 2), Range::all(), Range::all()), input);
-    normAssert(slice(out, Range::all(), Range(2, 4), Range::all(), Range::all()), input);
-
-    //
-
-    testLayerUsingCaffeModels("layer_concat_optim", DNN_TARGET_CPU, true, false);
-    testLayerUsingCaffeModels("layer_concat_shared_input", DNN_TARGET_CPU, true, false);
+    normAssert(slice(out, Range::all(), Range(0, 2), Range::all(), Range::all()), input, "", default_l1, default_lInf);
+    normAssert(slice(out, Range::all(), Range(2, 4), Range::all(), Range::all()), input, "", default_l1, default_lInf);
 }
 
 TEST_P(Test_Caffe_layers, Eltwise)
 {
-    testLayerUsingCaffeModels("layer_eltwise", GetParam());
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
+    testLayerUsingCaffeModels("layer_eltwise");
 }
 
 TEST_P(Test_Caffe_layers, PReLU)
 {
-    int targetId = GetParam();
-    testLayerUsingCaffeModels("layer_prelu", targetId, true);
-    testLayerUsingCaffeModels("layer_prelu_fc", targetId, true, false);
+    testLayerUsingCaffeModels("layer_prelu", true);
+}
+
+// TODO: fix an unstable test case
+TEST_P(Test_Caffe_layers, layer_prelu_fc)
+{
+    if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
+        throw SkipTestException("");
+    testLayerUsingCaffeModels("layer_prelu_fc", true, false);
 }
 
 //template<typename XMat>
@@ -311,13 +329,16 @@ TEST_P(Test_Caffe_layers, PReLU)
 //    );
 //}
 
-static void test_Reshape_Split_Slice_layers(int targetId)
+TEST_P(Test_Caffe_layers, Reshape_Split_Slice)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
+
     Net net = readNetFromCaffe(_tf("reshape_and_slice_routines.prototxt"));
     ASSERT_FALSE(net.empty());
 
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    net.setPreferableTarget(targetId);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
 
     Mat input(6, 12, CV_32F);
     RNG rng(0);
@@ -326,15 +347,10 @@ static void test_Reshape_Split_Slice_layers(int targetId)
     net.setInput(input, "input");
     Mat output = net.forward("output");
 
-    normAssert(input, output);
+    normAssert(input, output, "", default_l1, default_lInf);
 }
 
-TEST_P(Test_Caffe_layers, Reshape_Split_Slice)
-{
-    test_Reshape_Split_Slice_layers(GetParam());
-}
-
-TEST(Layer_Conv_Elu, Accuracy)
+TEST_P(Test_Caffe_layers, Conv_Elu)
 {
     Net net = readNetFromTensorflow(_tf("layer_elu_model.pb"));
     ASSERT_FALSE(net.empty());
@@ -343,10 +359,11 @@ TEST(Layer_Conv_Elu, Accuracy)
     Mat ref = blobFromNPY(_tf("layer_elu_out.npy"));
 
     net.setInput(inp, "input");
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
     Mat out = net.forward();
 
-    normAssert(ref, out);
+    normAssert(ref, out, "", default_l1, default_lInf);
 }
 
 class Layer_LSTM_Test : public ::testing::Test
@@ -496,37 +513,6 @@ TEST_F(Layer_RNN_Test, get_set_test)
     EXPECT_EQ(shape(outputs[1]), shape(nT, nS, nH));
 }
 
-void testLayerUsingDarknetModels(String basename, bool useDarknetModel = false, bool useCommonInputBlob = true)
-{
-    String cfg = _tf(basename + ".cfg");
-    String weights = _tf(basename + ".weights");
-
-    String inpfile = (useCommonInputBlob) ? _tf("blob.npy") : _tf(basename + ".input.npy");
-    String outfile = _tf(basename + ".npy");
-
-    Net net = readNetFromDarknet(cfg, (useDarknetModel) ? weights : String());
-    ASSERT_FALSE(net.empty());
-
-    Mat inp = blobFromNPY(inpfile);
-    Mat ref = blobFromNPY(outfile);
-
-    net.setInput(inp, "data");
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    Mat out = net.forward();
-
-    normAssert(ref, out);
-}
-
-TEST(Layer_Test_Region, Accuracy)
-{
-    testLayerUsingDarknetModels("region", false, false);
-}
-
-TEST(Layer_Test_Reorg, Accuracy)
-{
-    testLayerUsingDarknetModels("reorg", false, false);
-}
-
 TEST(Layer_Test_ROIPooling, Accuracy)
 {
     Net net = readNetFromCaffe(_tf("net_roi_pooling.prototxt"));
@@ -546,8 +532,10 @@ TEST(Layer_Test_ROIPooling, Accuracy)
 
 TEST_P(Test_Caffe_layers, FasterRCNN_Proposal)
 {
+    if ((backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ||
+        backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
     Net net = readNetFromCaffe(_tf("net_faster_rcnn_proposal.prototxt"));
-    net.setPreferableTarget(GetParam());
 
     Mat scores = blobFromNPY(_tf("net_faster_rcnn_proposal.scores.npy"));
     Mat deltas = blobFromNPY(_tf("net_faster_rcnn_proposal.deltas.npy"));
@@ -558,7 +546,8 @@ TEST_P(Test_Caffe_layers, FasterRCNN_Proposal)
     net.setInput(imInfo, "im_info");
 
     std::vector<Mat> outs;
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
     net.forward(outs, "output");
 
     for (int i = 0; i < 2; ++i)
@@ -573,7 +562,6 @@ TEST_P(Test_Caffe_layers, FasterRCNN_Proposal)
             EXPECT_EQ(countNonZero(outs[i].rowRange(numDets, outs[i].size[0])), 0);
     }
 }
-INSTANTIATE_TEST_CASE_P(/**/, Test_Caffe_layers, availableDnnTargets());
 
 typedef testing::TestWithParam<tuple<Vec4i, Vec2i, bool> > Scale_untrainable;
 TEST_P(Scale_untrainable, Accuracy)
@@ -739,8 +727,10 @@ INSTANTIATE_TEST_CASE_P(Layer_Test, Crop, Combine(
 
 // Check that by default average pooling layer should not count zero padded values
 // into the normalization area.
-TEST(Layer_Test_Average_pooling_kernel_area, Accuracy)
+TEST_P(Test_Caffe_layers, Average_pooling_kernel_area)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
     LayerParams lp;
     lp.name = "testAvePool";
     lp.type = "Pooling";
@@ -755,17 +745,21 @@ TEST(Layer_Test_Average_pooling_kernel_area, Accuracy)
     // ----+--
     // 7 8 | 9
     Mat inp = (Mat_<float>(3, 3) << 1, 2, 3, 4, 5, 6, 7, 8, 9);
-    Mat target = (Mat_<float>(2, 2) << (1 + 2 + 4 + 5) / 4.f, (3 + 6) / 2.f, (7 + 8) / 2.f, 9);
+    Mat ref = (Mat_<float>(2, 2) << (1 + 2 + 4 + 5) / 4.f, (3 + 6) / 2.f, (7 + 8) / 2.f, 9);
     Mat tmp = blobFromImage(inp);
     net.setInput(blobFromImage(inp));
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
     Mat out = net.forward();
-    normAssert(out, blobFromImage(target));
+    normAssert(out, blobFromImage(ref));
 }
 
 // Test PriorBoxLayer in case of no aspect ratios (just squared proposals).
-TEST(Layer_PriorBox, squares)
+TEST_P(Test_Caffe_layers, PriorBox_squares)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
+        (backend == DNN_BACKEND_OPENCV && (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16)))
+        throw SkipTestException("");
     LayerParams lp;
     lp.name = "testPriorBox";
     lp.type = "PriorBox";
@@ -783,14 +777,15 @@ TEST(Layer_PriorBox, squares)
     Mat inp(1, 2, CV_32F);
     randu(inp, -1, 1);
     net.setInput(blobFromImage(inp));
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
+    net.setPreferableBackend(backend);
+    net.setPreferableTarget(target);
     Mat out = net.forward();
 
-    Mat target = (Mat_<float>(4, 4) << 0.0, 0.0, 0.75, 1.0,
+    Mat ref = (Mat_<float>(4, 4) << 0.0, 0.0, 0.75, 1.0,
                                        0.25, 0.0, 1.0, 1.0,
                                        0.1f, 0.1f, 0.2f, 0.2f,
                                        0.1f, 0.1f, 0.2f, 0.2f);
-    normAssert(out.reshape(1, 4), target);
+    normAssert(out.reshape(1, 4), ref);
 }
 
 typedef TestWithParam<tuple<int, int> > Layer_Test_DWconv_Prelu;
@@ -1056,19 +1051,19 @@ TEST(Test_DLDT, multiple_networks)
 #endif  // HAVE_INF_ENGINE
 
 // Test a custom layer.
-class InterpLayer CV_FINAL : public Layer
+class CustomInterpLayer CV_FINAL : public Layer
 {
 public:
-    InterpLayer(const LayerParams &params) : Layer(params)
+    CustomInterpLayer(const LayerParams &params) : Layer(params)
     {
         zoomFactor = params.get<int>("zoom_factor", 0);
         outWidth = params.get<int>("width", 0);
         outHeight = params.get<int>("height", 0);
     }
 
-    static Ptr<InterpLayer> create(LayerParams& params)
+    static Ptr<Layer> create(LayerParams& params)
     {
-        return Ptr<InterpLayer>(new InterpLayer(params));
+        return Ptr<Layer>(new CustomInterpLayer(params));
     }
 
     virtual bool getMemoryShapes(const std::vector<std::vector<int> > &inputs,
@@ -1142,24 +1137,41 @@ public:
         }
     }
 
-    virtual void forward(InputArrayOfArrays, OutputArrayOfArrays, OutputArrayOfArrays) CV_OVERRIDE {}
+    void forward(InputArrayOfArrays inputs, OutputArrayOfArrays outputs, OutputArrayOfArrays internals) CV_OVERRIDE
+    {
+        CV_TRACE_FUNCTION();
+        CV_TRACE_ARG_VALUE(name, "name", name.c_str());
+
+        Layer::forward_fallback(inputs, outputs, internals);
+    }
 
 private:
     int outWidth, outHeight, zoomFactor;
 };
 
-TEST(Layer_Test_Interp_custom, Accuracy)
+TEST_P(Test_Caffe_layers, Interp)
 {
-    CV_DNN_REGISTER_LAYER_CLASS(Interp, InterpLayer);
-    testLayerUsingCaffeModels("layer_interp", DNN_TARGET_CPU, false, false);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
+    // Test a cusom layer.
+    CV_DNN_REGISTER_LAYER_CLASS(Interp, CustomInterpLayer);
+    try
+    {
+        testLayerUsingCaffeModels("layer_interp", false, false);
+    }
+    catch (...)
+    {
+        LayerFactory::unregisterLayer("Interp");
+        throw;
+    }
     LayerFactory::unregisterLayer("Interp");
-}
 
-TEST(Layer_Test_Interp, Accuracy)
-{
-    testLayerUsingCaffeModels("layer_interp", DNN_TARGET_CPU, false, false);
+    // Test an implemented layer.
+    testLayerUsingCaffeModels("layer_interp", false, false);
 }
 
+INSTANTIATE_TEST_CASE_P(/*nothing*/, Test_Caffe_layers, dnnBackendsAndTargets());
+
 TEST(Layer_Test_PoolingIndices, Accuracy)
 {
     Net net;
index 7ca1b98..16f1e5c 100644 (file)
@@ -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
index 4087822..ff1d0b3 100644 (file)
@@ -78,141 +78,170 @@ static std::string path(const std::string& file)
     return findDataFile("dnn/tensorflow/" + file, false);
 }
 
-static void runTensorFlowNet(const std::string& prefix, int targetId = DNN_TARGET_CPU, bool hasText = false,
-                             double l1 = 1e-5, double lInf = 1e-4,
-                             bool memoryLoad = false)
+class Test_TensorFlow_layers : public DNNTestLayer
 {
-    std::string netPath = path(prefix + "_net.pb");
-    std::string netConfig = (hasText ? path(prefix + "_net.pbtxt") : "");
-    std::string inpPath = path(prefix + "_in.npy");
-    std::string outPath = path(prefix + "_out.npy");
-
-    Net net;
-    if (memoryLoad)
+public:
+    void runTensorFlowNet(const std::string& prefix, bool hasText = false,
+                          double l1 = 0.0, double lInf = 0.0, bool memoryLoad = false)
     {
-        // Load files into a memory buffers
-        string dataModel;
-        ASSERT_TRUE(readFileInMemory(netPath, dataModel));
+        std::string netPath = path(prefix + "_net.pb");
+        std::string netConfig = (hasText ? path(prefix + "_net.pbtxt") : "");
+        std::string inpPath = path(prefix + "_in.npy");
+        std::string outPath = path(prefix + "_out.npy");
+
+        cv::Mat input = blobFromNPY(inpPath);
+        cv::Mat ref = blobFromNPY(outPath);
+        checkBackend(&input, &ref);
+
+        Net net;
+        if (memoryLoad)
+        {
+            // Load files into a memory buffers
+            string dataModel;
+            ASSERT_TRUE(readFileInMemory(netPath, dataModel));
+
+            string dataConfig;
+            if (hasText)
+                ASSERT_TRUE(readFileInMemory(netConfig, dataConfig));
+
+            net = readNetFromTensorflow(dataModel.c_str(), dataModel.size(),
+                                        dataConfig.c_str(), dataConfig.size());
+        }
+        else
+            net = readNetFromTensorflow(netPath, netConfig);
 
-        string dataConfig;
-        if (hasText)
-            ASSERT_TRUE(readFileInMemory(netConfig, dataConfig));
+        ASSERT_FALSE(net.empty());
 
-        net = readNetFromTensorflow(dataModel.c_str(), dataModel.size(),
-                                    dataConfig.c_str(), dataConfig.size());
+        net.setPreferableBackend(backend);
+        net.setPreferableTarget(target);
+        net.setInput(input);
+        cv::Mat output = net.forward();
+        normAssert(ref, output, "", l1 ? l1 : default_l1, lInf ? lInf : default_lInf);
     }
-    else
-        net = readNetFromTensorflow(netPath, netConfig);
-
-    ASSERT_FALSE(net.empty());
-
-    net.setPreferableBackend(DNN_BACKEND_OPENCV);
-    net.setPreferableTarget(targetId);
-
-    cv::Mat input = blobFromNPY(inpPath);
-    cv::Mat target = blobFromNPY(outPath);
-
-    net.setInput(input);
-    cv::Mat output = net.forward();
-    normAssert(target, output, "", l1, lInf);
-}
-
-typedef testing::TestWithParam<DNNTarget> Test_TensorFlow_layers;
+};
 
 TEST_P(Test_TensorFlow_layers, conv)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("single_conv", targetId);
-    runTensorFlowNet("atrous_conv2d_valid", targetId);
-    runTensorFlowNet("atrous_conv2d_same", targetId);
-    runTensorFlowNet("depthwise_conv2d", targetId);
-    runTensorFlowNet("keras_atrous_conv2d_same", targetId);
-    runTensorFlowNet("conv_pool_nchw", targetId);
+    runTensorFlowNet("single_conv");
+    runTensorFlowNet("atrous_conv2d_valid");
+    runTensorFlowNet("atrous_conv2d_same");
+    runTensorFlowNet("depthwise_conv2d");
+    runTensorFlowNet("keras_atrous_conv2d_same");
+    runTensorFlowNet("conv_pool_nchw");
 }
 
 TEST_P(Test_TensorFlow_layers, padding)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("padding_same", targetId);
-    runTensorFlowNet("padding_valid", targetId);
-    runTensorFlowNet("spatial_padding", targetId);
+    runTensorFlowNet("padding_same");
+    runTensorFlowNet("padding_valid");
+    runTensorFlowNet("spatial_padding");
 }
 
 TEST_P(Test_TensorFlow_layers, eltwise_add_mul)
 {
-    runTensorFlowNet("eltwise_add_mul", GetParam());
+    runTensorFlowNet("eltwise_add_mul");
+}
+
+TEST_P(Test_TensorFlow_layers, pad_and_concat)
+{
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
+    runTensorFlowNet("pad_and_concat");
 }
 
-TEST_P(Test_TensorFlow_layers, concat)
+TEST_P(Test_TensorFlow_layers, concat_axis_1)
 {
-    runTensorFlowNet("pad_and_concat", GetParam());
-    runTensorFlowNet("concat_axis_1", GetParam());
+    runTensorFlowNet("concat_axis_1");
 }
 
 TEST_P(Test_TensorFlow_layers, batch_norm)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("batch_norm", targetId);
-    runTensorFlowNet("fused_batch_norm", targetId);
-    runTensorFlowNet("batch_norm_text", targetId, true);
-    runTensorFlowNet("mvn_batch_norm", targetId);
-    runTensorFlowNet("mvn_batch_norm_1x1", targetId);
-    runTensorFlowNet("unfused_batch_norm", targetId);
-    runTensorFlowNet("fused_batch_norm_no_gamma", targetId);
-    runTensorFlowNet("unfused_batch_norm_no_gamma", targetId);
+    runTensorFlowNet("batch_norm");
+    runTensorFlowNet("batch_norm", false, 0.0, 0.0, true);
+    runTensorFlowNet("fused_batch_norm");
+    runTensorFlowNet("fused_batch_norm", false, 0.0, 0.0, true);
+    runTensorFlowNet("batch_norm_text", true);
+    runTensorFlowNet("batch_norm_text", true, 0.0, 0.0, true);
+    runTensorFlowNet("unfused_batch_norm");
+    runTensorFlowNet("fused_batch_norm_no_gamma");
+    runTensorFlowNet("unfused_batch_norm_no_gamma");
+}
+
+TEST_P(Test_TensorFlow_layers, mvn_batch_norm)
+{
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
+    runTensorFlowNet("mvn_batch_norm");
+    runTensorFlowNet("mvn_batch_norm_1x1");
 }
 
 TEST_P(Test_TensorFlow_layers, pooling)
 {
-    int targetId = GetParam();
-    cv::ocl::Device d = cv::ocl::Device::getDefault();
-    bool loosenFlag = targetId == DNN_TARGET_OPENCL && d.isIntel() && d.type() == cv::ocl::Device::TYPE_CPU;
-    runTensorFlowNet("max_pool_even", targetId);
-    runTensorFlowNet("max_pool_odd_valid", targetId);
-    runTensorFlowNet("ave_pool_same", targetId);
-    runTensorFlowNet("max_pool_odd_same", targetId, false, loosenFlag ? 3e-5 : 1e-5, loosenFlag ? 3e-4 : 1e-4);
-    runTensorFlowNet("reduce_mean", targetId);  // an average pooling over all spatial dimensions.
+    runTensorFlowNet("max_pool_even");
+    runTensorFlowNet("max_pool_odd_valid");
+    runTensorFlowNet("max_pool_odd_same");
+    runTensorFlowNet("reduce_mean");  // an average pooling over all spatial dimensions.
+}
+
+// TODO: fix tests and replace to pooling
+TEST_P(Test_TensorFlow_layers, ave_pool_same)
+{
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
+    runTensorFlowNet("ave_pool_same");
 }
 
 TEST_P(Test_TensorFlow_layers, deconvolution)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("deconvolution", targetId);
-    runTensorFlowNet("deconvolution_same", targetId);
-    runTensorFlowNet("deconvolution_stride_2_same", targetId);
-    runTensorFlowNet("deconvolution_adj_pad_valid", targetId);
-    runTensorFlowNet("deconvolution_adj_pad_same", targetId);
-    runTensorFlowNet("keras_deconv_valid", targetId);
-    runTensorFlowNet("keras_deconv_same", targetId);
+    runTensorFlowNet("deconvolution");
+    runTensorFlowNet("deconvolution_same");
+    runTensorFlowNet("deconvolution_stride_2_same");
+    runTensorFlowNet("deconvolution_adj_pad_valid");
+    runTensorFlowNet("deconvolution_adj_pad_same");
+    runTensorFlowNet("keras_deconv_valid");
+    runTensorFlowNet("keras_deconv_same");
 }
 
 TEST_P(Test_TensorFlow_layers, matmul)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("matmul", targetId);
-    runTensorFlowNet("nhwc_reshape_matmul", targetId);
-    runTensorFlowNet("nhwc_transpose_reshape_matmul", targetId);
+    if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
+        throw SkipTestException("");
+    runTensorFlowNet("matmul");
+    runTensorFlowNet("nhwc_reshape_matmul");
+    runTensorFlowNet("nhwc_transpose_reshape_matmul");
 }
 
 TEST_P(Test_TensorFlow_layers, reshape)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("shift_reshape_no_reorder", targetId);
-    runTensorFlowNet("reshape_no_reorder", targetId);
-    runTensorFlowNet("reshape_reduce", targetId);
-    runTensorFlowNet("flatten", targetId, true);
-    runTensorFlowNet("unfused_flatten", targetId);
-    runTensorFlowNet("unfused_flatten_unknown_batch", targetId);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
+    runTensorFlowNet("shift_reshape_no_reorder");
+    runTensorFlowNet("reshape_no_reorder");
+    runTensorFlowNet("reshape_reduce");
+}
+
+TEST_P(Test_TensorFlow_layers, flatten)
+{
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE &&
+        (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
+    runTensorFlowNet("flatten", true);
+    runTensorFlowNet("unfused_flatten");
+    runTensorFlowNet("unfused_flatten_unknown_batch");
 }
 
 TEST_P(Test_TensorFlow_layers, l2_normalize)
 {
-    int targetId = GetParam();
-    runTensorFlowNet("l2_normalize", targetId);
-    runTensorFlowNet("l2_normalize_3d", targetId);
+    runTensorFlowNet("l2_normalize");
 }
 
-INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_layers, availableDnnTargets());
+// TODO: fix it and add to l2_normalize
+TEST_P(Test_TensorFlow_layers, l2_normalize_3d)
+{
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
+    runTensorFlowNet("l2_normalize_3d");
+}
 
 typedef testing::TestWithParam<DNNTarget> Test_TensorFlow_nets;
 
@@ -359,90 +388,95 @@ TEST_P(Test_TensorFlow_nets, EAST_text_detection)
 
 INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_nets, availableDnnTargets());
 
-typedef testing::TestWithParam<DNNTarget> Test_TensorFlow_fp16;
-
-TEST_P(Test_TensorFlow_fp16, tests)
+TEST_P(Test_TensorFlow_layers, fp16_weights)
 {
-    int targetId = GetParam();
-    const float l1 = 7e-4;
-    const float lInf = 1e-2;
-    runTensorFlowNet("fp16_single_conv", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_deconvolution", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_max_pool_odd_same", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_padding_valid", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_eltwise_add_mul", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_max_pool_odd_valid", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_pad_and_concat", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_max_pool_even", targetId, false, l1, lInf);
-    runTensorFlowNet("fp16_padding_same", targetId, false, l1, lInf);
+    const float l1 = 0.00071;
+    const float lInf = 0.012;
+    runTensorFlowNet("fp16_single_conv", false, l1, lInf);
+    runTensorFlowNet("fp16_deconvolution", false, l1, lInf);
+    runTensorFlowNet("fp16_max_pool_odd_same", false, l1, lInf);
+    runTensorFlowNet("fp16_padding_valid", false, l1, lInf);
+    runTensorFlowNet("fp16_eltwise_add_mul", false, l1, lInf);
+    runTensorFlowNet("fp16_max_pool_odd_valid", false, l1, lInf);
+    runTensorFlowNet("fp16_max_pool_even", false, l1, lInf);
+    runTensorFlowNet("fp16_padding_same", false, l1, lInf);
 }
 
-INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_fp16,
-                        Values(DNN_TARGET_CPU, DNN_TARGET_OPENCL, DNN_TARGET_OPENCL_FP16));
+// TODO: fix pad_and_concat and add this test case to fp16_weights
+TEST_P(Test_TensorFlow_layers, fp16_pad_and_concat)
+{
+    const float l1 = 0.00071;
+    const float lInf = 0.012;
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
+    runTensorFlowNet("fp16_pad_and_concat", false, l1, lInf);
+}
 
-TEST(Test_TensorFlow, defun)
+TEST_P(Test_TensorFlow_layers, defun)
 {
     runTensorFlowNet("defun_dropout");
 }
 
-TEST(Test_TensorFlow, quantized)
+TEST_P(Test_TensorFlow_layers, quantized)
 {
     runTensorFlowNet("uint8_single_conv");
 }
 
-TEST(Test_TensorFlow, lstm)
+TEST_P(Test_TensorFlow_layers, lstm)
 {
-    runTensorFlowNet("lstm", DNN_TARGET_CPU, true);
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE ||
+        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
+    runTensorFlowNet("lstm", true);
+    runTensorFlowNet("lstm", true, 0.0, 0.0, true);
 }
 
-TEST(Test_TensorFlow, split)
+TEST_P(Test_TensorFlow_layers, split)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        throw SkipTestException("");
     runTensorFlowNet("split_equals");
 }
 
-TEST(Test_TensorFlow, resize_nearest_neighbor)
+TEST_P(Test_TensorFlow_layers, resize_nearest_neighbor)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_MYRIAD)
+        throw SkipTestException("");
     runTensorFlowNet("resize_nearest_neighbor");
 }
 
-TEST(Test_TensorFlow, slice)
+TEST_P(Test_TensorFlow_layers, slice)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE &&
+        (target == DNN_TARGET_OPENCL || target == DNN_TARGET_OPENCL_FP16))
+        throw SkipTestException("");
     runTensorFlowNet("slice_4d");
 }
 
-TEST(Test_TensorFlow, softmax)
+TEST_P(Test_TensorFlow_layers, softmax)
 {
     runTensorFlowNet("keras_softmax");
 }
 
-TEST(Test_TensorFlow, relu6)
+TEST_P(Test_TensorFlow_layers, relu6)
 {
     runTensorFlowNet("keras_relu6");
-    runTensorFlowNet("keras_relu6", DNN_TARGET_CPU, /*hasText*/ true);
+    runTensorFlowNet("keras_relu6", /*hasText*/ true);
 }
 
-TEST(Test_TensorFlow, keras_mobilenet_head)
+TEST_P(Test_TensorFlow_layers, keras_mobilenet_head)
 {
     runTensorFlowNet("keras_mobilenet_head");
 }
 
-TEST(Test_TensorFlow, memory_read)
-{
-    double l1 = 1e-5;
-    double lInf = 1e-4;
-    runTensorFlowNet("lstm", DNN_TARGET_CPU, true, l1, lInf, true);
-
-    runTensorFlowNet("batch_norm", DNN_TARGET_CPU, false, l1, lInf, true);
-    runTensorFlowNet("fused_batch_norm", DNN_TARGET_CPU, false, l1, lInf, true);
-    runTensorFlowNet("batch_norm_text", DNN_TARGET_CPU, true, l1, lInf, true);
-}
-
-TEST(Test_TensorFlow, resize_bilinear)
+TEST_P(Test_TensorFlow_layers, resize_bilinear)
 {
     runTensorFlowNet("resize_bilinear");
     runTensorFlowNet("resize_bilinear_factor");
 }
 
+INSTANTIATE_TEST_CASE_P(/**/, Test_TensorFlow_layers, dnnBackendsAndTargets());
+
 TEST(Test_TensorFlow, two_inputs)
 {
     Net net = readNet(path("two_inputs_net.pbtxt"));
index 5fe3fe1..c1abdc9 100644 (file)
@@ -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.