Merge pull request #16010 from YashasSamaga:cuda4dnn-fp16-tests

* enable tests for DNN_TARGET_CUDA_FP16

* disable deconvolution tests

* disable shortcut tests

* fix typos and some minor changes

* dnn(test): skip CUDA FP16 test too (run_pool_max)

diff --git a/modules/dnn/test/test_backends.cpp b/modules/dnn/test/test_backends.cpp
index 2e8c9ec..a5297c7 100644 (file)
--- a/modules/dnn/test/test_backends.cpp
+++ b/modules/dnn/test/test_backends.cpp
@@ -168,6 +168,8 @@ TEST_P(DNNTestNetwork, ENet)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL_FP16);
+    if (backend == DNN_BACKEND_CUDA && target == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
     processNet("dnn/Enet-model-best.net", "", Size(512, 512), "l367_Deconvolution",
                target == DNN_TARGET_OPENCL ? "dnn/halide_scheduler_opencl_enet.yml" :
                                              "dnn/halide_scheduler_enet.yml",
@@ -182,11 +184,11 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_Caffe)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_HALIDE);
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f / 127.5, Size(300, 300), Scalar(127.5, 127.5, 127.5), false);
-    float diffScores = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1.5e-2 : 0.0;
-    float diffSquares = (target == DNN_TARGET_MYRIAD) ? 0.063  : 0.0;
+    float scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1.5e-2 : 0.0;
+    float iouDiff = (target == DNN_TARGET_MYRIAD) ? 0.063  : 0.0;
     float detectionConfThresh = (target == DNN_TARGET_MYRIAD) ? 0.252  : FLT_MIN;
          processNet("dnn/MobileNetSSD_deploy.caffemodel", "dnn/MobileNetSSD_deploy.prototxt",
-                    inp, "detection_out", "", diffScores, diffSquares, detectionConfThresh);
+                    inp, "detection_out", "", scoreDiff, iouDiff, detectionConfThresh);
     expectNoFallbacksFromIE(net);
 }
 
@@ -201,10 +203,19 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_Caffe_Different_Width_Height)
 #endif
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f / 127.5, Size(300, 560), Scalar(127.5, 127.5, 127.5), false);
-    float diffScores  = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.029 : 0.0;
-    float diffSquares = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.09  : 0.0;
+    float scoreDiff = 0.0, iouDiff = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.029;
+        iouDiff = 0.09;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.03;
+        iouDiff = 0.08;
+    }
     processNet("dnn/MobileNetSSD_deploy.caffemodel", "dnn/MobileNetSSD_deploy.prototxt",
-                inp, "detection_out", "", diffScores, diffSquares);
+                inp, "detection_out", "", scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
 
@@ -216,11 +227,20 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_v1_TensorFlow)
 
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f, Size(300, 300), Scalar(), false);
-    float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.095 : 0.0;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.09 : 0.0;
     float detectionConfThresh = (target == DNN_TARGET_MYRIAD) ? 0.216 : 0.2;
+    float scoreDiff = 0.0, iouDiff = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.095;
+        iouDiff = 0.09;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.007;
+        iouDiff = 0.08;
+    }
     processNet("dnn/ssd_mobilenet_v1_coco_2017_11_17.pb", "dnn/ssd_mobilenet_v1_coco_2017_11_17.pbtxt",
-               inp, "detection_out", "", l1, lInf, detectionConfThresh);
+               inp, "detection_out", "", scoreDiff, iouDiff, detectionConfThresh);
     expectNoFallbacksFromIE(net);
 }
 
@@ -240,10 +260,19 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_v1_TensorFlow_Different_Width_Height)
 
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f, Size(300, 560), Scalar(), false);
-    float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.012 : 0.0;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.06 : 0.0;
+    float scoreDiff = 0.0, iouDiff = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.012;
+        iouDiff = 0.06;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.007;
+        iouDiff = 0.06;
+    }
     processNet("dnn/ssd_mobilenet_v1_coco_2017_11_17.pb", "dnn/ssd_mobilenet_v1_coco_2017_11_17.pbtxt",
-               inp, "detection_out", "", l1, lInf);
+               inp, "detection_out", "", scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
 
@@ -255,10 +284,19 @@ TEST_P(DNNTestNetwork, MobileNet_SSD_v2_TensorFlow)
 
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f, Size(300, 300), Scalar(), false);
-    float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.013 : 2e-5;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.062 : 0.0;
+    float scoreDiff = 2e-5, iouDiff = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.013;
+        iouDiff = 0.062;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.02;
+        iouDiff = 0.07;
+    }
     processNet("dnn/ssd_mobilenet_v2_coco_2018_03_29.pb", "dnn/ssd_mobilenet_v2_coco_2018_03_29.pbtxt",
-               inp, "detection_out", "", l1, lInf, 0.25);
+               inp, "detection_out", "", scoreDiff, iouDiff, 0.25);
     expectNoFallbacksFromIE(net);
 }
 
@@ -268,12 +306,25 @@ TEST_P(DNNTestNetwork, SSD_VGG16)
                  CV_TEST_TAG_DEBUG_VERYLONG);
     if (backend == DNN_BACKEND_HALIDE && target == DNN_TARGET_CPU)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_HALIDE);  // TODO HALIDE_CPU
-    double scoreThreshold = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.0325 : 0.0;
-    const float lInf = (target == DNN_TARGET_MYRIAD) ? 0.032 : 0.0;
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f, Size(300, 300), Scalar(), false);
+    float scoreDiff = 0.0, iouDiff = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        scoreDiff = 0.0325;
+    }
+    else if (target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.0325;
+        iouDiff = 0.032;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.03;
+    }
+
     processNet("dnn/VGG_ILSVRC2016_SSD_300x300_iter_440000.caffemodel",
-               "dnn/ssd_vgg16.prototxt", inp, "detection_out", "", scoreThreshold, lInf);
+               "dnn/ssd_vgg16.prototxt", inp, "detection_out", "", scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
 
@@ -384,10 +435,19 @@ TEST_P(DNNTestNetwork, Inception_v2_SSD_TensorFlow)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_HALIDE);
     Mat sample = imread(findDataFile("dnn/street.png"));
     Mat inp = blobFromImage(sample, 1.0f, Size(300, 300), Scalar(), false);
-    float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.015 : 0.0;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.0731 : 0.0;
+    float scoreDiff = 0.0, iouDiff = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.015;
+        iouDiff = 0.0731;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.015;
+        iouDiff = 0.08;
+    }
     processNet("dnn/ssd_inception_v2_coco_2017_11_17.pb", "dnn/ssd_inception_v2_coco_2017_11_17.pbtxt",
-               inp, "detection_out", "", l1, lInf);
+               inp, "detection_out", "", scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
 
@@ -400,11 +460,18 @@ TEST_P(DNNTestNetwork, DenseNet_121)
     float l1 = 0.0, lInf = 0.0;
     if (target == DNN_TARGET_OPENCL_FP16)
     {
-        l1 = 2e-2; lInf = 9e-2;
+        l1 = 2e-2;
+        lInf = 9e-2;
     }
     else if (target == DNN_TARGET_MYRIAD)
     {
-        l1 = 0.1; lInf = 0.6;
+        l1 = 0.1;
+        lInf = 0.6;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.008;
+        lInf = 0.05;
     }
     processNet("dnn/DenseNet_121.caffemodel", "dnn/DenseNet_121.prototxt", Size(224, 224), "", "", l1, lInf);
     if (target != DNN_TARGET_MYRIAD || getInferenceEngineVPUType() != CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
@@ -431,8 +498,17 @@ TEST_P(DNNTestNetwork, FastNeuralStyle_eccv16)
     Mat img = imread(findDataFile("dnn/googlenet_1.png"));
     Mat inp = blobFromImage(img, 1.0, Size(320, 240), Scalar(103.939, 116.779, 123.68), false, false);
     // Output image has values in range [-143.526, 148.539].
-    float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.4 : 4e-5;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 7.45 : 2e-3;
+    float l1 = 4e-5, lInf = 2e-3;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 0.4;
+        lInf = 7.45;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.3;
+        lInf = 7.2;
+    }
     processNet("dnn/fast_neural_style_eccv16_starry_night.t7", "", inp, "", "", l1, lInf);
 #if defined(HAVE_INF_ENGINE) && INF_ENGINE_VER_MAJOR_GE(2019010000)
     expectNoFallbacksFromIE(net);

diff --git a/modules/dnn/test/test_caffe_importer.cpp b/modules/dnn/test/test_caffe_importer.cpp
index 0607c7d..d0996db 100644 (file)
--- a/modules/dnn/test/test_caffe_importer.cpp
+++ b/modules/dnn/test/test_caffe_importer.cpp
@@ -150,8 +150,17 @@ TEST_P(Test_Caffe_nets, Axpy)
             }
         }
     }
-    float l1 = (target == DNN_TARGET_OPENCL_FP16) ? 2e-4 : 1e-5;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16) ? 1e-3 : 1e-4;
+    float l1 = 1e-5, lInf = 1e-4;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        l1 = 2e-4;
+        lInf = 1e-3;
+    }
+    else if(target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.0002;
+        lInf = 0.0007;
+    }
     normAssert(ref, out, "", l1, lInf);
 }
 
@@ -287,8 +296,17 @@ TEST_P(Reproducibility_MobileNet_SSD, Accuracy)
 
     ASSERT_EQ(out.size[2], 100);
 
-    const float scores_diff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 1.5e-2 : 1e-5;
-    const float boxes_iou_diff = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 6.3e-2 : 1e-4;
+    float scores_diff = 1e-5, boxes_iou_diff = 1e-4;
+    if (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD)
+    {
+        scores_diff = 1.5e-2;
+        boxes_iou_diff = 6.3e-2;
+    }
+    else if (targetId == DNN_TARGET_CUDA_FP16)
+    {
+        scores_diff = 0.015;
+        boxes_iou_diff = 0.07;
+    }
     Mat ref = blobFromNPY(_tf("mobilenet_ssd_caffe_out.npy"));
     normAssertDetections(ref, out, "", FLT_MIN, scores_diff, boxes_iou_diff);
 
@@ -477,11 +495,21 @@ TEST_P(Test_Caffe_nets, Colorization)
     Mat out = net.forward();
 
     // Reference output values are in range [-29.1, 69.5]
-    double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.25 : 4e-4;
-    double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 5.3 : 3e-3;
-    if (target == DNN_TARGET_MYRIAD && getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+    double l1 = 4e-4, lInf = 3e-3;
+    if (target == DNN_TARGET_OPENCL_FP16)
     {
-        l1 = 0.5; lInf = 11;
+        l1 = 0.25;
+        lInf = 5.3;
+    }
+    else if (target == DNN_TARGET_MYRIAD)
+    {
+        l1 = (getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X) ? 0.5 : 0.25;
+        lInf = (getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X) ? 11 : 5.3;
+    }
+    else if(target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.21;
+        lInf = 4.5;
     }
     normAssert(out, ref, "", l1, lInf);
     expectNoFallbacksFromIE(net);
@@ -518,6 +546,10 @@ TEST_P(Test_Caffe_nets, DenseNet_121)
     {
         l1 = 0.11; lInf = 0.5;
     }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.04; lInf = 0.2;
+    }
     normAssert(outs[0], ref, "", l1, lInf);
     if (target != DNN_TARGET_MYRIAD || getInferenceEngineVPUType() != CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
         expectNoFallbacksFromIE(model);
@@ -663,6 +695,8 @@ TEST_P(Test_Caffe_nets, FasterRCNN_zf)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD);
+    if (target == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
     static Mat ref = (Mat_<float>(3, 7) << 0, 2, 0.90121, 120.407, 115.83, 570.586, 528.395,
                                            0, 7, 0.988779, 469.849, 75.1756, 718.64, 186.762,
                                            0, 12, 0.967198, 138.588, 206.843, 329.766, 553.176);
@@ -680,8 +714,17 @@ TEST_P(Test_Caffe_nets, RFCN)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16);
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD);
-    double scoreDiff = (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ? 4e-3 : default_l1;
-    double iouDiff = (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ? 8e-2 : default_lInf;
+    float scoreDiff = default_l1, iouDiff = default_lInf;
+    if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
+    {
+        scoreDiff = 4e-3;
+        iouDiff = 8e-2;
+    }
+    if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.0034;
+        iouDiff = 0.11;
+    }
     static Mat ref = (Mat_<float>(2, 7) << 0, 7, 0.991359, 491.822, 81.1668, 702.573, 178.234,
                                            0, 12, 0.94786, 132.093, 223.903, 338.077, 566.16);
     testFaster("rfcn_pascal_voc_resnet50.prototxt", "resnet50_rfcn_final.caffemodel", ref, scoreDiff, iouDiff);

diff --git a/modules/dnn/test/test_common.impl.hpp b/modules/dnn/test/test_common.impl.hpp
index 8721b64..16114d5 100644 (file)
--- a/modules/dnn/test/test_common.impl.hpp
+++ b/modules/dnn/test/test_common.impl.hpp
@@ -239,9 +239,8 @@ testing::internal::ParamGenerator< tuple<Backend, Target> > dnnBackendsAndTarget
 #ifdef HAVE_CUDA
     if(withCUDA)
     {
-        //for (auto target : getAvailableTargets(DNN_BACKEND_CUDA))
-        //    targets.push_back(make_tuple(DNN_BACKEND_CUDA, target));
-        targets.push_back(make_tuple(DNN_BACKEND_CUDA, DNN_TARGET_CUDA));
+        for (auto target : getAvailableTargets(DNN_BACKEND_CUDA))
+            targets.push_back(make_tuple(DNN_BACKEND_CUDA, target));
     }
 #endif
 

diff --git a/modules/dnn/test/test_darknet_importer.cpp b/modules/dnn/test/test_darknet_importer.cpp
index eced695..2a60659 100644 (file)
--- a/modules/dnn/test/test_darknet_importer.cpp
+++ b/modules/dnn/test/test_darknet_importer.cpp
@@ -320,9 +320,18 @@ TEST_P(Test_Darknet_nets, YoloVoc)
                                     1, 6,  0.667770f, 0.446555f, 0.453578f, 0.499986f, 0.519167f,  // a car
                                     1, 6,  0.844947f, 0.637058f, 0.460398f, 0.828508f, 0.66427f);  // a car
 
-    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) ? 0.018 : 3e-4;
     double nmsThreshold = (target == DNN_TARGET_MYRIAD) ? 0.397 : 0.4;
+    double scoreDiff = 8e-5, iouDiff = 3e-4;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 1e-2;
+        iouDiff = 0.018;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.03;
+        iouDiff = 0.018;
+    }
 
     std::string config_file = "yolo-voc.cfg";
     std::string weights_file = "yolo-voc.weights";
@@ -353,8 +362,17 @@ TEST_P(Test_Darknet_nets, TinyYoloVoc)
                                     1, 6,  0.651450f, 0.460526f, 0.458019f, 0.522527f, 0.5341f,    // a car
                                     1, 6,  0.928758f, 0.651024f, 0.463539f, 0.823784f, 0.654998f); // a car
 
-    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) ? 0.018 : 3e-4;
+    double scoreDiff = 8e-5, iouDiff = 3e-4;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 8e-3;
+        iouDiff = 0.018;
+    }
+    else if(target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.008;
+        iouDiff = 0.02;
+    }
 
     std::string config_file = "tiny-yolo-voc.cfg";
     std::string weights_file = "tiny-yolo-voc.weights";
@@ -453,9 +471,17 @@ TEST_P(Test_Darknet_nets, YOLOv3)
                                     1, 2,  0.989633f, 0.450719f, 0.463353f, 0.496305f, 0.522258f,  // a car
                                     1, 2,  0.997412f, 0.647584f, 0.459939f, 0.821038f, 0.663947f); // a car
 
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.006 : 8e-5;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.018 : 3e-4;
-
+    double scoreDiff = 8e-5, iouDiff = 3e-4;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.006;
+        iouDiff = 0.018;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.04;
+        iouDiff = 0.03;
+    }
     std::string config_file = "yolov3.cfg";
     std::string weights_file = "yolov3.weights";
 
@@ -501,6 +527,8 @@ INSTANTIATE_TEST_CASE_P(/**/, Test_Darknet_nets, dnnBackendsAndTargets());
 
 TEST_P(Test_Darknet_layers, shortcut)
 {
+    if (backend == DNN_BACKEND_CUDA)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA);
     testDarknetLayer("shortcut");
     testDarknetLayer("shortcut_leaky");
     testDarknetLayer("shortcut_unequal");

diff --git a/modules/dnn/test/test_halide_layers.cpp b/modules/dnn/test/test_halide_layers.cpp
index 11668b4..a68dd19 100644 (file)
--- a/modules/dnn/test/test_halide_layers.cpp
+++ b/modules/dnn/test/test_halide_layers.cpp
@@ -16,7 +16,7 @@ using namespace cv;
 using namespace cv::dnn;
 using namespace testing;
 
-static void test(Mat& input, Net& net, Backend backendId, Target targetId, bool skipCheck = false, bool randInput = true)
+static void test(Mat& input, Net& net, Backend backendId, Target targetId, bool skipCheck = false, bool randInput = true, double l1 = 0.0, double lInf = 0.0)
 {
     DNNTestLayer::checkBackend(backendId, targetId);
     if (randInput)
@@ -33,8 +33,12 @@ static void test(Mat& input, Net& net, Backend backendId, Target targetId, bool
     if (skipCheck)
         return;
 
-    double l1, lInf;
-    DNNTestLayer::getDefaultThresholds(backendId, targetId, &l1, &lInf);
+    double default_l1, default_lInf;
+    DNNTestLayer::getDefaultThresholds(backendId, targetId, &default_l1, &default_lInf);
+    if (l1 == 0.0)
+        l1 = default_l1;
+    if (lInf == 0.0)
+        lInf = default_lInf;
 #if 0
     std::cout << "l1=" << l1 << "  lInf=" << lInf << std::endl;
     std::cout << outputDefault.reshape(1, outputDefault.total()).t() << std::endl;
@@ -43,11 +47,11 @@ static void test(Mat& input, Net& net, Backend backendId, Target targetId, bool
     normAssert(outputDefault, outputHalide, "", l1, lInf);
 }
 
-static void test(LayerParams& params, Mat& input, Backend backendId, Target targetId, bool skipCheck = false)
+static void test(LayerParams& params, Mat& input, Backend backendId, Target targetId, bool skipCheck = false, double l1 = 0.0, double lInf = 0.0)
 {
     Net net;
     net.addLayerToPrev(params.name, params.type, params);
-    test(input, net, backendId, targetId, skipCheck);
+    test(input, net, backendId, targetId, skipCheck, true, l1, lInf);
 }
 
 static inline testing::internal::ParamGenerator<tuple<Backend, Target> > dnnBackendsAndTargetsWithHalide()
@@ -174,6 +178,9 @@ TEST_P(Deconvolution, Accuracy)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_X);
 #endif
 
+    if (targetId == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
+
     int sz[] = {inChannels, outChannels / group, kernel.height, kernel.width};
     Mat weights(4, &sz[0], CV_32F);
     randu(weights, -1.0f, 1.0f);
@@ -414,7 +421,11 @@ TEST_P(FullyConnected, Accuracy)
 
     int sz[] = {1, inChannels, inSize.height, inSize.width};
     Mat input(4, &sz[0], CV_32F);
-    test(lp, input, backendId, targetId);
+
+    double l1 = 0.0;
+    if (targetId == DNN_TARGET_CUDA_FP16)
+        l1 = 0.015;
+    test(lp, input, backendId, targetId, false, true, l1);
 }
 
 INSTANTIATE_TEST_CASE_P(Layer_Test_Halide, FullyConnected, Combine(
@@ -497,7 +508,7 @@ TEST_P(Test_Halide_layers, MaxPoolUnpool)
 ////////////////////////////////////////////////////////////////////////////////
 static const int kNumChannels = 3;
 
-void testInPlaceActivation(LayerParams& lp, Backend backendId, Target targetId)
+void testInPlaceActivation(LayerParams& lp, Backend backendId, Target targetId, double l1 = 0.0, double lInf = 0.0)
 {
     EXPECT_FALSE(lp.name.empty());
 
@@ -517,7 +528,7 @@ void testInPlaceActivation(LayerParams& lp, Backend backendId, Target targetId)
 
     int sz[] = {1, kNumChannels, 10, 10};
     Mat input(4, &sz[0], CV_32F);
-    test(input, net, backendId, targetId);
+    test(input, net, backendId, targetId, false, true, l1, lInf);
 }
 
 typedef TestWithParam<tuple<bool, bool, float, tuple<Backend, Target> > > BatchNorm;

diff --git a/modules/dnn/test/test_layers.cpp b/modules/dnn/test/test_layers.cpp
index e4ac578..b3fa22f 100644 (file)
--- a/modules/dnn/test/test_layers.cpp
+++ b/modules/dnn/test/test_layers.cpp
@@ -141,6 +141,8 @@ TEST_P(Test_Caffe_layers, Convolution)
 
 TEST_P(Test_Caffe_layers, DeConvolution)
 {
+    if(target == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
     testLayerUsingCaffeModels("layer_deconvolution", true, false);
 }
 
@@ -372,7 +374,13 @@ TEST_P(Test_Caffe_layers, Conv_Elu)
     net.setPreferableTarget(target);
     Mat out = net.forward();
 
-    normAssert(ref, out, "", default_l1, default_lInf);
+    double l1 = default_l1, lInf = default_lInf;
+    if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.0002;
+        lInf = 0.0005;
+    }
+    normAssert(ref, out, "", l1, lInf);
 }
 
 class Layer_LSTM_Test : public ::testing::Test
@@ -843,6 +851,11 @@ TEST_P(Test_Caffe_layers, PriorBox_repeated)
 
     double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1e-3 : 1e-5;
     double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1e-3 : 1e-4;
+    if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 7e-5;
+        lInf = 0.0005;
+    }
     normAssert(out, ref, "", l1, lInf);
 }
 
@@ -876,7 +889,9 @@ TEST_P(Test_Caffe_layers, PriorBox_squares)
                                        0.25, 0.0, 1.0, 1.0,
                                        0.1f, 0.1f, 0.2f, 0.2f,
                                        0.1f, 0.1f, 0.2f, 0.2f);
-    double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 2e-5 : 1e-5;
+    double l1 = 1e-5;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD || target == DNN_TARGET_CUDA_FP16)
+        l1 = 2e-5;
     normAssert(out.reshape(1, 4), ref, "", l1);
 }
 
@@ -1225,6 +1240,11 @@ TEST_P(Test_DLDT_two_inputs, as_backend)
     // Output values are in range [0, 637.5].
     double l1 = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.06 : 1e-6;
     double lInf = (targetId == DNN_TARGET_OPENCL_FP16 || targetId == DNN_TARGET_MYRIAD) ? 0.3 : 1e-5;
+    if (targetId == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.06;
+        lInf = 0.3;
+    }
     normAssert(out, ref, "", l1, lInf);
 }
 
@@ -1537,8 +1557,17 @@ TEST_P(Layer_Test_ShuffleChannel, Accuracy)
     net.setPreferableTarget(targetId);
     Mat out = net.forward();
 
-    double l1 = (targetId == DNN_TARGET_OPENCL_FP16) ? 5e-2 : 1e-5;
-    double lInf = (targetId == DNN_TARGET_OPENCL_FP16) ? 7e-2 : 1e-4;
+    double l1 = 1e-5, lInf = 1e-4;
+    if (targetId == DNN_TARGET_OPENCL_FP16)
+    {
+        l1 = 5e-2;
+        lInf = 7e-2;
+    }
+    else if (targetId == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.06;
+        lInf = 0.07;
+    }
     for (int n = 0; n < inpShapeVec[0]; ++n)
     {
         for (int c = 0; c < inpShapeVec[1]; ++c)
@@ -1593,6 +1622,9 @@ TEST_P(Layer_Test_Eltwise_unequal, accuracy_input_0_truncate)
     int backendId = get<0>(get<1>(GetParam()));
     int targetId = get<1>(get<1>(GetParam()));
 
+    if (backendId == DNN_BACKEND_CUDA)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA);
+
     Net net;
     LayerParams lp;
     lp.type = "Eltwise";
@@ -1656,6 +1688,9 @@ TEST_P(Layer_Test_Eltwise_unequal, accuracy_input_0)
     int backendId = get<0>(get<1>(GetParam()));
     int targetId = get<1>(get<1>(GetParam()));
 
+    if (backendId == DNN_BACKEND_CUDA)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA);
+
     Net net;
     LayerParams lp;
     lp.type = "Eltwise";

diff --git a/modules/dnn/test/test_model.cpp b/modules/dnn/test/test_model.cpp
index 7a4de4e..bbe4ce4 100644 (file)
--- a/modules/dnn/test/test_model.cpp
+++ b/modules/dnn/test/test_model.cpp
@@ -157,9 +157,13 @@ TEST_P(Test_Model, DetectRegion)
     bool swapRB = true;
 
     double confThreshold = 0.24;
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1e-2 : 8e-5;
-    double iouDiff = (target == DNN_TARGET_MYRIAD || target == DNN_TARGET_OPENCL_FP16) ? 1.6e-2 : 1e-5;
     double nmsThreshold = (target == DNN_TARGET_MYRIAD) ? 0.397 : 0.4;
+    double scoreDiff = 8e-5, iouDiff = 1e-5;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD || target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 1e-2;
+        iouDiff = 1.6e-2;
+    }
 
     testDetectModel(weights_file, config_file, img_path, refClassIds, refConfidences,
                     refBoxes, scoreDiff, iouDiff, confThreshold, nmsThreshold, size,
@@ -188,11 +192,15 @@ TEST_P(Test_Model, DetectionOutput)
     Scalar mean = Scalar(102.9801, 115.9465, 122.7717);
     Size size{800, 600};
 
-    double scoreDiff = (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ?
-                        4e-3 : default_l1;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16) ? 1.8e-1 : 1e-5;
+    double scoreDiff = default_l1, iouDiff = 1e-5;
     float confThreshold = 0.8;
     double nmsThreshold = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_CUDA_FP16)
+    {
+        if (backend == DNN_BACKEND_OPENCV)
+            scoreDiff = 4e-3;
+        iouDiff = 1.8e-1;
+    }
 
     testDetectModel(weights_file, config_file, img_path, refClassIds, refConfidences, refBoxes,
                     scoreDiff, iouDiff, confThreshold, nmsThreshold, size, mean);
@@ -232,10 +240,22 @@ TEST_P(Test_Model, DetectionMobilenetSSD)
     double scale = 1.0 / 127.5;
     Size size{300, 300};
 
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 1.7e-2 : 1e-5;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || (target == DNN_TARGET_MYRIAD &&
-                      getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)) ? 6.91e-2 : 1e-5;
-
+    double scoreDiff = 1e-5, iouDiff = 1e-5;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        scoreDiff = 1.7e-2;
+        iouDiff = 6.91e-2;
+    }
+    else if (target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 1.7e-2;
+        if (getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
+            iouDiff = 6.91e-2;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 4e-4;
+    }
     float confThreshold = FLT_MIN;
     double nmsThreshold = 0.0;
 
@@ -263,6 +283,10 @@ TEST_P(Test_Model, Keypoints_pose)
     Scalar mean = Scalar(128, 128, 128);
     bool swapRB = false;
 
+    // Ref. Range: [58.6875, 508.625]
+    if (target == DNN_TARGET_CUDA_FP16)
+        norm = 20; // l1 = 1.5, lInf = 20
+
     testKeypointsModel(weights, "", inp, exp, norm, size, mean, scale, swapRB);
 }
 
@@ -283,8 +307,11 @@ TEST_P(Test_Model, Keypoints_face)
     Scalar mean = Scalar();
     bool swapRB = false;
 
-    testKeypointsModel(weights, "", inp, exp, norm, size, mean, scale, swapRB);
+    // Ref. Range: [-1.1784188, 1.7758257]
+    if (target == DNN_TARGET_CUDA_FP16)
+        norm = 0.004; // l1 = 0.0006, lInf = 0.004
 
+    testKeypointsModel(weights, "", inp, exp, norm, size, mean, scale, swapRB);
 }
 
 TEST_P(Test_Model, Detection_normalized)
@@ -301,10 +328,14 @@ TEST_P(Test_Model, Detection_normalized)
     double scale = 1.0 / 127.5;
     Size size{300, 300};
 
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 5e-3 : 1e-5;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.09 : 1e-5;
+    double scoreDiff = 1e-5, iouDiff = 1e-5;
     float confThreshold = FLT_MIN;
     double nmsThreshold = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD || target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 5e-3;
+        iouDiff = 0.09;
+    }
     testDetectModel(weights_file, config_file, img_path, refClassIds, refConfidences, refBoxes,
                     scoreDiff, iouDiff, confThreshold, nmsThreshold, size, mean, scale);
 }

diff --git a/modules/dnn/test/test_onnx_importer.cpp b/modules/dnn/test/test_onnx_importer.cpp
index ce8a43a..7f4a18c 100644 (file)
--- a/modules/dnn/test/test_onnx_importer.cpp
+++ b/modules/dnn/test/test_onnx_importer.cpp
@@ -590,8 +590,17 @@ TEST_P(Test_ONNX_nets, TinyYolov2)
 #endif
 
     // output range: [-11; 8]
-    double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.017 : default_l1;
-    double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.14 : default_lInf;
+    double l1 =  default_l1, lInf = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 0.017;
+        lInf = 0.14;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.018;
+        lInf = 0.16;
+    }
     testONNXModels("tiny_yolo2", pb, l1, lInf);
 }
 
@@ -620,17 +629,23 @@ TEST_P(Test_ONNX_nets, LResNet100E_IR)
         if (target == DNN_TARGET_MYRIAD)      applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
     }
 
-    double l1 = default_l1;
-    double lInf = default_lInf;
+    double l1 = default_l1, lInf = default_lInf;
     // output range: [-3; 3]
-    if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) {
+    if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
+    {
         l1 = 0.009;
         lInf = 0.035;
     }
-    else if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_CPU) {
+    else if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_CPU)
+    {
         l1 = 4.6e-5;
         lInf = 1.9e-4;
     }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.008;
+        lInf = 0.04;
+    }
     testONNXModels("LResNet100E_IR", pb, l1, lInf);
 }
 
@@ -747,8 +762,12 @@ TEST_P(Test_ONNX_nets, Resnet34_kinetics)
     net.setPreferableTarget(target);
 
     // output range [-5, 11]
-    float l1 = 0.0013;
-    float lInf = 0.009;
+    float l1 = 0.0013, lInf = 0.009;
+    if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.008;
+        lInf = 0.04;
+    }
 
     checkBackend(&input0, &ref0);
     net.setInput(input0);

diff --git a/modules/dnn/test/test_tf_importer.cpp b/modules/dnn/test/test_tf_importer.cpp
index c49ed51..f563e25 100644 (file)
--- a/modules/dnn/test/test_tf_importer.cpp
+++ b/modules/dnn/test/test_tf_importer.cpp
@@ -225,8 +225,17 @@ TEST_P(Test_TensorFlow_layers, slim_batch_norm)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NGRAPH)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_NGRAPH);
     // Output values range: [-40.0597, 207.827]
-    double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.041 : default_l1;
-    double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.33 : default_lInf;
+    double l1 =  default_l1, lInf = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 0.041;
+        lInf = 0.33;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.005;
+        lInf = 0.33;
+    }
     runTensorFlowNet("slim_batch_norm", false, l1, lInf);
 }
 
@@ -300,9 +309,8 @@ TEST_P(Test_TensorFlow_layers, AvePooling3D)
 
 TEST_P(Test_TensorFlow_layers, deconvolution)
 {
-    if(backend == DNN_BACKEND_CUDA)
-        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA); /* bugged */
-
+    if (backend == DNN_BACKEND_CUDA)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA);
     runTensorFlowNet("deconvolution");
     runTensorFlowNet("deconvolution_same");
     runTensorFlowNet("deconvolution_stride_2_same");
@@ -428,8 +436,16 @@ TEST_P(Test_TensorFlow_nets, MobileNet_SSD)
     net.setInput(inp);
     Mat out = net.forward();
 
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.0043 : default_l1;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.037 : default_lInf;
+    double scoreDiff = default_l1, iouDiff = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.0043;
+        iouDiff = 0.037;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        iouDiff = 0.04;
+    }
     normAssertDetections(ref, out, "", 0.2, scoreDiff, iouDiff);
 #if defined(INF_ENGINE_RELEASE) && INF_ENGINE_RELEASE >= 2019010000
     expectNoFallbacksFromIE(net);
@@ -466,8 +482,17 @@ TEST_P(Test_TensorFlow_nets, Inception_v2_SSD)
                                     0, 10, 0.95932811, 0.38349164, 0.32528657, 0.40387636, 0.39165527,
                                     0, 10, 0.93973452, 0.66561931, 0.37841269, 0.68074018, 0.42907384);
 
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.0097 : default_l1;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.09 : default_lInf;
+    double scoreDiff = default_l1, iouDiff = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.0097;
+        iouDiff = 0.09;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 6e-3;
+        iouDiff = 0.05;
+    }
     normAssertDetections(ref, out, "", 0.5, scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
@@ -489,10 +514,18 @@ TEST_P(Test_TensorFlow_nets, MobileNet_v1_SSD)
     Mat out = net.forward();
 
     Mat ref = blobFromNPY(findDataFile("dnn/tensorflow/ssd_mobilenet_v1_coco_2017_11_17.detection_out.npy"));
-    float scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.011 : 1.5e-5;
-    float iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.012 : 1e-3;
+    float scoreDiff = 1.5e-5, iouDiff = 1e-3;
     float detectionConfThresh = (target == DNN_TARGET_MYRIAD) ? 0.35 : 0.3;
-
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.011;
+        iouDiff = 0.012;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.006;
+        iouDiff = 0.01;
+    }
 #if defined(INF_ENGINE_RELEASE)
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD &&
         getInferenceEngineVPUType() == CV_DNN_INFERENCE_ENGINE_VPU_TYPE_MYRIAD_X)
@@ -530,6 +563,9 @@ TEST_P(Test_TensorFlow_nets, Faster_RCNN)
     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL_FP16);
 
+    if (backend == DNN_BACKEND_CUDA && target == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
+
     checkBackend();
 
     double scoresDiff = backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 ? 2.9e-5 : 1e-5;
@@ -574,8 +610,17 @@ TEST_P(Test_TensorFlow_nets, MobileNet_v1_SSD_PPN)
     net.setInput(blob);
     Mat out = net.forward();
 
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.048 : 1.1e-5;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.058 : default_lInf;
+    double scoreDiff = 1.1e-5, iouDiff = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 0.048;
+        iouDiff = 0.058;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 0.006;
+        iouDiff = 0.05;
+    }
     normAssertDetections(ref, out, "", 0.45, scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
@@ -604,8 +649,17 @@ TEST_P(Test_TensorFlow_nets, opencv_face_detector_uint8)
                                     0, 1, 0.98977017, 0.23901358, 0.09084064, 0.29902688, 0.1769477,
                                     0, 1, 0.97203469, 0.67965847, 0.06876482, 0.73999709, 0.1513494,
                                     0, 1, 0.95097077, 0.51901293, 0.45863652, 0.5777427, 0.5347801);
-    double scoreDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 4e-3 : 3.4e-3;
-    double iouDiff = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.024 : 1e-2;
+    double scoreDiff = 3.4e-3, iouDiff = 1e-2;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        scoreDiff = 4e-3;
+        iouDiff = 0.024;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        scoreDiff = 4e-3;
+        iouDiff = 0.02;
+    }
     normAssertDetections(ref, out, "", 0.9, scoreDiff, iouDiff);
     expectNoFallbacksFromIE(net);
 }
@@ -673,6 +727,11 @@ TEST_P(Test_TensorFlow_nets, EAST_text_detection)
         lInf_scores = 0.41;
         l1_geometry = 0.28; lInf_geometry = 5.94;
     }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        lInf_scores = 0.1;
+        l1_geometry = 0.3; lInf_geometry = 7;
+    }
     else
     {
         l1_geometry = 1e-4, lInf_geometry = 3e-3;
@@ -695,7 +754,8 @@ TEST_P(Test_TensorFlow_layers, fp16_weights)
     runTensorFlowNet("fp16_padding_valid", false, l1, lInf);
     // Reference output values are in range [0.0889, 1.651]
     runTensorFlowNet("fp16_max_pool_even", false, (target == DNN_TARGET_MYRIAD) ? 0.003 : l1, lInf);
-    if (target == DNN_TARGET_MYRIAD) {
+    if (target == DNN_TARGET_MYRIAD)
+    {
         l1 = 0.0041;
         lInf = 0.024;
     }

diff --git a/modules/dnn/test/test_torch_importer.cpp b/modules/dnn/test/test_torch_importer.cpp
index 5343fae..1f4bc1f 100644 (file)
--- a/modules/dnn/test/test_torch_importer.cpp
+++ b/modules/dnn/test/test_torch_importer.cpp
@@ -112,8 +112,17 @@ public:
 TEST_P(Test_Torch_layers, run_convolution)
 {
     // Output reference values are in range [23.4018, 72.0181]
-    double l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.08 : default_l1;
-    double lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 0.42 : default_lInf;
+    double l1 = default_l1, lInf = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 0.08;
+        lInf = 0.42;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.08;
+        lInf = 0.5;
+    }
     runTorchNet("net_conv", "", false, true, true, l1, lInf);
 }
 
@@ -121,7 +130,10 @@ TEST_P(Test_Torch_layers, run_pool_max)
 {
     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_OPENCL_FP16);
-    runTorchNet("net_pool_max", "", true);
+    if (target == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
+    double l1 = 0.0, lInf = 0.0;
+    runTorchNet("net_pool_max", "", true, false, true, l1, lInf);
 }
 
 TEST_P(Test_Torch_layers, run_pool_ave)
@@ -145,9 +157,17 @@ TEST_P(Test_Torch_layers, run_reshape)
 TEST_P(Test_Torch_layers, run_reshape_single_sample)
 {
     // Reference output values in range [14.4586, 18.4492].
-    runTorchNet("net_reshape_single_sample", "", false, false, true,
-                (target == DNN_TARGET_MYRIAD || target == DNN_TARGET_OPENCL_FP16) ? 0.033 : default_l1,
-                (target == DNN_TARGET_MYRIAD || target == DNN_TARGET_OPENCL_FP16) ? 0.05 : default_lInf);
+    double l1 = default_l1, lInf = default_lInf;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 0.033;
+        lInf = 0.05;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.01;
+    }
+    runTorchNet("net_reshape_single_sample", "", false, false, true, l1, lInf);
 }
 
 TEST_P(Test_Torch_layers, run_linear)
@@ -164,8 +184,16 @@ TEST_P(Test_Torch_layers, run_concat)
 
 TEST_P(Test_Torch_layers, run_depth_concat)
 {
-    runTorchNet("net_depth_concat", "", false, true, true, 0.0,
-                target == DNN_TARGET_OPENCL_FP16 ? 0.021 : 0.0);
+    double lInf = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        lInf = 0.021;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        lInf = 0.03;
+    }
+    runTorchNet("net_depth_concat", "", false, true, true, 0.0, lInf);
 }
 
 TEST_P(Test_Torch_layers, run_deconv)
@@ -211,9 +239,18 @@ TEST_P(Test_Torch_layers, net_conv_gemm_lrn)
 {
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target == DNN_TARGET_MYRIAD)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
-    runTorchNet("net_conv_gemm_lrn", "", false, true, true,
-                target == DNN_TARGET_OPENCL_FP16 ? 0.046 : 0.0,
-                target == DNN_TARGET_OPENCL_FP16 ? 0.023 : 0.0);
+    double l1 = 0.0, lInf = 0.0;
+    if (target == DNN_TARGET_OPENCL_FP16)
+    {
+        l1 = 0.046;
+        lInf = 0.023;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.0042;
+        lInf = 0.021;
+    }
+    runTorchNet("net_conv_gemm_lrn", "", false, true, true, l1, lInf);
 }
 
 TEST_P(Test_Torch_layers, net_inception_block)
@@ -291,8 +328,17 @@ TEST_P(Test_Torch_nets, OpenFace_accuracy)
 
     // Reference output values are in range [-0.17212, 0.263492]
     // on Myriad problem layer: l4_Pooling - does not use pads_begin
-    float l1 = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 2e-3 : 1e-5;
-    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 5e-3 : 1e-3;
+    float l1 = 1e-5, lInf = 1e-3;
+    if (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD)
+    {
+        l1 = 2e-3;
+        lInf = 5e-3;
+    }
+    else if (target == DNN_TARGET_CUDA_FP16)
+    {
+        l1 = 0.0004;
+        lInf = 0.0012;
+    }
     Mat outRef = readTorchBlob(_tf("net_openface_output.dat"), true);
     normAssert(out, outRef, "", l1, lInf);
 }
@@ -343,6 +389,8 @@ TEST_P(Test_Torch_nets, ENet_accuracy)
     checkBackend();
     if (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16)
         throw SkipTestException("");
+    if (backend == DNN_BACKEND_CUDA && target == DNN_TARGET_CUDA_FP16)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_CUDA_FP16);
     if (backend == DNN_BACKEND_INFERENCE_ENGINE_NN_BUILDER_2019 && target != DNN_TARGET_CPU)
     {
         if (target == DNN_TARGET_OPENCL_FP16) applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_OPENCL_FP16, CV_TEST_TAG_DNN_SKIP_IE_NN_BUILDER);
@@ -448,6 +496,10 @@ TEST_P(Test_Torch_nets, FastNeuralStyle_accuracy)
             else
                 EXPECT_LE(normL1, 0.6f);
         }
+        else if(target == DNN_TARGET_CUDA_FP16)
+        {
+            normAssert(out, refBlob, "", 0.6, 25);
+        }
         else
             normAssert(out, refBlob, "", 0.5, 1.1);
     }