Enable some tests for clDNN plugin from Intel's Inference Engine

cmake/OpenCVDetectInferenceEngine.cmake

@@ -15,6 +15,10 @@ macro(ie_fail)
     return()
 endmacro()
 
+if(NOT HAVE_CXX11)
+    ie_fail()
+endif()
+
 if(NOT INF_ENGINE_ROOT_DIR OR NOT EXISTS "${INF_ENGINE_ROOT_DIR}/include/inference_engine.hpp")
     set(ie_root_paths "${INF_ENGINE_ROOT_DIR}")
     if(DEFINED ENV{INTEL_CVSDK_DIR})

modules/dnn/perf/perf_net.cpp

@@ -95,24 +95,18 @@ PERF_TEST_P_(DNNTestNetwork, AlexNet)
 
 PERF_TEST_P_(DNNTestNetwork, GoogLeNet)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
-        throw SkipTestException("");
     processNet("dnn/bvlc_googlenet.caffemodel", "dnn/bvlc_googlenet.prototxt",
             "", Mat(cv::Size(224, 224), CV_32FC3));
 }
 
 PERF_TEST_P_(DNNTestNetwork, ResNet_50)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
-        throw SkipTestException("");
     processNet("dnn/ResNet-50-model.caffemodel", "dnn/ResNet-50-deploy.prototxt",
             "resnet_50.yml", Mat(cv::Size(224, 224), CV_32FC3));
 }
 
 PERF_TEST_P_(DNNTestNetwork, SqueezeNet_v1_1)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
-        throw SkipTestException("");
     processNet("dnn/squeezenet_v1.1.caffemodel", "dnn/squeezenet_v1.1.prototxt",
             "squeezenet_v1_1.yml", Mat(cv::Size(227, 227), CV_32FC3));
 }

modules/dnn/src/dnn.cpp

@@ -1255,6 +1255,15 @@ struct Net::Impl
                     if (weightableLayer->_biases)
                         weightableLayer->_biases = convertFp16(weightableLayer->_biases);
                 }
+                else
+                {
+                    for (const auto& weights : {"weights", "biases"})
+                    {
+                        auto it = ieNode->layer->blobs.find(weights);
+                        if (it != ieNode->layer->blobs.end())
+                            it->second = convertFp16(it->second);
+                    }
+                }
             }
 
             ieNode->connect(ld.inputBlobsWrappers, ld.outputBlobsWrappers);

modules/dnn/src/layers/prior_box_layer.cpp

@@ -295,6 +295,19 @@ public:
         return false;
     }
 
+    void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs) CV_OVERRIDE
+    {
+        CV_Assert(inputs.size() > 1, inputs[0]->dims == 4, inputs[1]->dims == 4);
+        int layerWidth = inputs[0]->size[3];
+        int layerHeight = inputs[0]->size[2];
+
+        int imageWidth = inputs[1]->size[3];
+        int imageHeight = inputs[1]->size[2];
+
+        _stepY = _stepY == 0 ? (static_cast<float>(imageHeight) / layerHeight) : _stepY;
+        _stepX = _stepX == 0 ? (static_cast<float>(imageWidth) / layerWidth) : _stepX;
+    }
+
 #ifdef HAVE_OPENCL
     bool forward_ocl(InputArrayOfArrays inps, OutputArrayOfArrays outs, OutputArrayOfArrays internals)
     {
@@ -310,16 +323,6 @@ public:
         int _imageWidth = inputs[1].size[3];
         int _imageHeight = inputs[1].size[2];
 
-        float stepX, stepY;
-        if (_stepX == 0 || _stepY == 0)
-        {
-            stepX = static_cast<float>(_imageWidth) / _layerWidth;
-            stepY = static_cast<float>(_imageHeight) / _layerHeight;
-        } else {
-            stepX = _stepX;
-            stepY = _stepY;
-        }
-
         if (umat_offsetsX.empty())
         {
             Mat offsetsX(1, _offsetsX.size(), CV_32FC1, &_offsetsX[0]);
@@ -339,8 +342,8 @@ public:
 
         ocl::Kernel kernel("prior_box", ocl::dnn::prior_box_oclsrc);
         kernel.set(0, (int)nthreads);
-        kernel.set(1, (float)stepX);
+        kernel.set(1, (float)_stepX);
-        kernel.set(2, (float)stepY);
+        kernel.set(2, (float)_stepY);
         kernel.set(3, ocl::KernelArg::PtrReadOnly(umat_offsetsX));
         kernel.set(4, ocl::KernelArg::PtrReadOnly(umat_offsetsY));
         kernel.set(5, (int)_offsetsX.size());
@@ -410,15 +413,6 @@ public:
         int _imageWidth = inputs[1]->size[3];
         int _imageHeight = inputs[1]->size[2];
 
-        float stepX, stepY;
-        if (_stepX == 0 || _stepY == 0) {
-          stepX = static_cast<float>(_imageWidth) / _layerWidth;
-          stepY = static_cast<float>(_imageHeight) / _layerHeight;
-        } else {
-          stepX = _stepX;
-          stepY = _stepY;
-        }
-
         float* outputPtr = outputs[0].ptr<float>();
         float _boxWidth, _boxHeight;
         for (size_t h = 0; h < _layerHeight; ++h)
@@ -431,8 +425,8 @@ public:
                     _boxHeight = _boxHeights[i];
                     for (int j = 0; j < _offsetsX.size(); ++j)
                     {
-                        float center_x = (w + _offsetsX[j]) * stepX;
+                        float center_x = (w + _offsetsX[j]) * _stepX;
-                        float center_y = (h + _offsetsY[j]) * stepY;
+                        float center_y = (h + _offsetsY[j]) * _stepY;
                         outputPtr = addPrior(center_x, center_y, _boxWidth, _boxHeight, _imageWidth,
                                              _imageHeight, _bboxesNormalized, outputPtr);
                     }
@@ -495,7 +489,7 @@ public:
                 ieLayer->params["aspect_ratio"] += format(",%f", _aspectRatios[i]);
         }
 
-        ieLayer->params["flip"] = _flip ? "1" : "0";
+        ieLayer->params["flip"] = "0";  // We already flipped aspect ratios.
         ieLayer->params["clip"] = _clip ? "1" : "0";
 
         CV_Assert(!_variance.empty());
@@ -503,12 +497,20 @@ public:
         for (int i = 1; i < _variance.size(); ++i)
             ieLayer->params["variance"] += format(",%f", _variance[i]);
 
-        ieLayer->params["step"] = _stepX == _stepY ? format("%f", _stepX) : "0";
+        if (_stepX == _stepY)
-        ieLayer->params["step_h"] = _stepY;
+        {
-        ieLayer->params["step_w"] = _stepX;
+            ieLayer->params["step"] = format("%f", _stepX);
-
+            ieLayer->params["step_h"] = "0.0";
+            ieLayer->params["step_w"] = "0.0";
+        }
+        else
+        {
+            ieLayer->params["step"] = "0.0";
+            ieLayer->params["step_h"] = format("%f", _stepY);
+            ieLayer->params["step_w"] = format("%f", _stepX);
+        }
         CV_Assert(_offsetsX.size() == 1, _offsetsY.size() == 1, _offsetsX[0] == _offsetsY[0]);
-        ieLayer->params["offset"] = format("%f", _offsetsX[0]);;
+        ieLayer->params["offset"] = format("%f", _offsetsX[0]);
 
         return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
 #endif  // HAVE_INF_ENGINE

modules/dnn/src/op_inf_engine.cpp

@@ -233,8 +233,16 @@ InferenceEngine::StatusCode
 InfEngineBackendNet::getLayerByName(const char *layerName, InferenceEngine::CNNLayerPtr &out,
                                     InferenceEngine::ResponseDesc *resp) noexcept
 {
-    CV_Error(Error::StsNotImplemented, "");
+    for (auto& l : layers)
-    return InferenceEngine::StatusCode::OK;
+    {
+        if (l->name == layerName)
+        {
+            out = l;
+            return InferenceEngine::StatusCode::OK;
+        }
+    }
+    CV_Error(Error::StsObjectNotFound, cv::format("Cannot find a layer %s", layerName));
+    return InferenceEngine::StatusCode::NOT_FOUND;
 }
 
 void InfEngineBackendNet::setTargetDevice(InferenceEngine::TargetDevice device) noexcept

modules/dnn/test/test_backends.cpp

@@ -23,9 +23,9 @@ public:
     }
 
     void processNet(const std::string& weights, const std::string& proto,
-                    Size inpSize, const std::string& outputLayer,
+                    Size inpSize, const std::string& outputLayer = "",
                     const std::string& halideScheduler = "",
-                    double l1 = 1e-5, double lInf = 1e-4)
+                    double l1 = 0.0, double lInf = 0.0)
     {
         // Create a common input blob.
         int blobSize[] = {1, 3, inpSize.height, inpSize.width};
@@ -36,9 +36,9 @@ public:
     }
 
     void processNet(std::string weights, std::string proto,
-                    Mat inp, const std::string& outputLayer,
+                    Mat inp, const std::string& outputLayer = "",
                     std::string halideScheduler = "",
-                    double l1 = 1e-5, double lInf = 1e-4)
+                    double l1 = 0.0, double lInf = 0.0)
     {
         if (backend == DNN_BACKEND_DEFAULT && target == DNN_TARGET_OPENCL)
         {
@@ -49,6 +49,16 @@ public:
                 throw SkipTestException("OpenCL is not available/disabled in OpenCV");
             }
         }
+        if (target == DNN_TARGET_OPENCL_FP16)
+        {
+            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;
+        }
         weights = findDataFile(weights, false);
         if (!proto.empty())
             proto = findDataFile(proto, false);
@@ -71,31 +81,28 @@ public:
         Mat out = net.forward(outputLayer).clone();
 
         if (outputLayer == "detection_out")
-            normAssertDetections(outDefault, out, "First run", 0.2);
+            normAssertDetections(outDefault, out, "First run", 0.2, l1, lInf);
         else
             normAssert(outDefault, out, "First run", l1, lInf);
 
         // Test 2: change input.
-        inp *= 0.1f;
+        float* inpData = (float*)inp.data;
+        for (int i = 0; i < inp.size[0] * inp.size[1]; ++i)
+        {
+            Mat slice(inp.size[2], inp.size[3], CV_32F, inpData);
+            cv::flip(slice, slice, 1);
+            inpData += slice.total();
+        }
         netDefault.setInput(inp);
         net.setInput(inp);
         outDefault = netDefault.forward(outputLayer).clone();
         out = net.forward(outputLayer).clone();
 
         if (outputLayer == "detection_out")
-            checkDetections(outDefault, out, "Second run", l1, lInf);
+            normAssertDetections(outDefault, out, "Second run", 0.2, l1, lInf);
         else
             normAssert(outDefault, out, "Second run", l1, lInf);
     }
-
-    void checkDetections(const Mat& out, const Mat& ref, const std::string& msg,
-                         float l1, float lInf, int top = 5)
-    {
-        top = std::min(std::min(top, out.size[2]), out.size[3]);
-        std::vector<cv::Range> range(4, cv::Range::all());
-        range[2] = cv::Range(0, top);
-        normAssert(out(range), ref(range));
-    }
 };
 
 TEST_P(DNNTestNetwork, AlexNet)
@@ -110,8 +117,6 @@ TEST_P(DNNTestNetwork, AlexNet)
 
 TEST_P(DNNTestNetwork, ResNet_50)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
-        throw SkipTestException("");
     processNet("dnn/ResNet-50-model.caffemodel", "dnn/ResNet-50-deploy.prototxt",
                Size(224, 224), "prob",
                target == DNN_TARGET_OPENCL ? "dnn/halide_scheduler_opencl_resnet_50.yml" :
@@ -120,8 +125,6 @@ TEST_P(DNNTestNetwork, ResNet_50)
 
 TEST_P(DNNTestNetwork, SqueezeNet_v1_1)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
-        throw SkipTestException("");
     processNet("dnn/squeezenet_v1.1.caffemodel", "dnn/squeezenet_v1.1.prototxt",
                Size(227, 227), "prob",
                target == DNN_TARGET_OPENCL ? "dnn/halide_scheduler_opencl_squeezenet_v1_1.yml" :
@@ -130,8 +133,6 @@ TEST_P(DNNTestNetwork, SqueezeNet_v1_1)
 
 TEST_P(DNNTestNetwork, GoogLeNet)
 {
-    if (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16)
-        throw SkipTestException("");
     processNet("dnn/bvlc_googlenet.caffemodel", "dnn/bvlc_googlenet.prototxt",
                Size(224, 224), "prob");
 }
@@ -180,7 +181,7 @@ TEST_P(DNNTestNetwork, SSD_VGG16)
 {
     if (backend == DNN_BACKEND_DEFAULT && target == DNN_TARGET_OPENCL ||
         backend == DNN_BACKEND_HALIDE && target == DNN_TARGET_CPU ||
-        backend == DNN_BACKEND_INFERENCE_ENGINE)
+        backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_CPU)
         throw SkipTestException("");
     processNet("dnn/VGG_ILSVRC2016_SSD_300x300_iter_440000.caffemodel",
                "dnn/ssd_vgg16.prototxt", Size(300, 300), "detection_out");
@@ -189,30 +190,24 @@ TEST_P(DNNTestNetwork, SSD_VGG16)
 TEST_P(DNNTestNetwork, OpenPose_pose_coco)
 {
     if (backend == DNN_BACKEND_HALIDE) throw SkipTestException("");
-    double l1 = target == DNN_TARGET_OPENCL_FP16 ? 3e-5 : 1e-5;
-    double lInf = target == DNN_TARGET_OPENCL_FP16 ? 3e-3 : 1e-4;
     processNet("dnn/openpose_pose_coco.caffemodel", "dnn/openpose_pose_coco.prototxt",
-               Size(368, 368), "", "", l1, lInf);
+               Size(368, 368));
 }
 
 TEST_P(DNNTestNetwork, OpenPose_pose_mpi)
 {
     if (backend == DNN_BACKEND_HALIDE) throw SkipTestException("");
-    double l1 = target == DNN_TARGET_OPENCL_FP16 ? 4e-5 : 1e-5;
-    double lInf = target == DNN_TARGET_OPENCL_FP16 ? 7e-3 : 1e-4;
     processNet("dnn/openpose_pose_mpi.caffemodel", "dnn/openpose_pose_mpi.prototxt",
-               Size(368, 368), "", "", l1, lInf);
+               Size(368, 368));
 }
 
 TEST_P(DNNTestNetwork, OpenPose_pose_mpi_faster_4_stages)
 {
     if (backend == DNN_BACKEND_HALIDE) throw SkipTestException("");
-    double l1 = target == DNN_TARGET_OPENCL_FP16 ? 5e-5 : 1e-5;
-    double lInf = target == DNN_TARGET_OPENCL_FP16 ? 5e-3 : 1e-4;
     // The same .caffemodel but modified .prototxt
     // See https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/src/openpose/pose/poseParameters.cpp
     processNet("dnn/openpose_pose_mpi.caffemodel", "dnn/openpose_pose_mpi_faster_4_stages.prototxt",
-               Size(368, 368), "", "", l1, lInf);
+               Size(368, 368));
 }
 
 TEST_P(DNNTestNetwork, OpenFace)