Merge pull request #11739 from dkurt:more_ie_models

modules/dnn/perf/perf_net.cpp

@@ -144,7 +144,8 @@ PERF_TEST_P_(DNNTestNetwork, SSD)
 PERF_TEST_P_(DNNTestNetwork, OpenFace)
 {
     if (backend == DNN_BACKEND_HALIDE ||
-        backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_CPU)
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
         throw SkipTestException("");
     processNet("dnn/openface_nn4.small2.v1.t7", "", "",
             Mat(cv::Size(96, 96), CV_32FC3));
@@ -248,6 +249,15 @@ PERF_TEST_P_(DNNTestNetwork, EAST_text_detection)
     processNet("dnn/frozen_east_text_detection.pb", "", "", Mat(cv::Size(320, 320), CV_32FC3));
 }
 
+PERF_TEST_P_(DNNTestNetwork, FastNeuralStyle_eccv16)
+{
+    if (backend == DNN_BACKEND_HALIDE ||
+        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
+        throw SkipTestException("");
+    processNet("dnn/fast_neural_style_eccv16_starry_night.t7", "", "", Mat(cv::Size(320, 240), CV_32FC3));
+}
+
 const tuple<DNNBackend, DNNTarget> testCases[] = {
 #ifdef HAVE_HALIDE
     tuple<DNNBackend, DNNTarget>(DNN_BACKEND_HALIDE, DNN_TARGET_CPU),

modules/dnn/src/layers/convolution_layer.cpp

@@ -81,9 +81,10 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
-        return backendId == DNN_BACKEND_OPENCV ||
-               backendId == DNN_BACKEND_HALIDE && haveHalide() ||
-               backendId == DNN_BACKEND_INFERENCE_ENGINE && haveInfEngine();
+        if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
+            return preferableTarget != DNN_TARGET_MYRIAD || type != "Deconvolution" || adjustPad == Size();
+        else
+            return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
     }
 
     void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs) CV_OVERRIDE

modules/dnn/src/layers/elementwise_layers.cpp

@@ -115,9 +115,7 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
-        return backendId == DNN_BACKEND_OPENCV ||
-               backendId == DNN_BACKEND_HALIDE && haveHalide() ||
-               backendId == DNN_BACKEND_INFERENCE_ENGINE && haveInfEngine();
+        return func.supportBackend(backendId, this->preferableTarget);
     }
 
     virtual Ptr<BackendNode> tryAttach(const Ptr<BackendNode>& node) CV_OVERRIDE
@@ -238,6 +236,12 @@ struct ReLUFunctor
 
     explicit ReLUFunctor(float slope_=1.f) : slope(slope_) {}
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         float s = slope;
@@ -353,6 +357,12 @@ struct ReLU6Functor
         CV_Assert(minValue <= maxValue);
     }
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
@@ -445,6 +455,12 @@ struct TanHFunctor
 {
     typedef TanHLayer Layer;
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
@@ -509,6 +525,12 @@ struct SigmoidFunctor
 {
     typedef SigmoidLayer Layer;
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
@@ -575,6 +597,11 @@ struct ELUFunctor
 
     explicit ELUFunctor() {}
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
@@ -638,6 +665,11 @@ struct AbsValFunctor
 {
     typedef AbsLayer Layer;
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
@@ -701,6 +733,11 @@ struct BNLLFunctor
 {
     typedef BNLLLayer Layer;
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
@@ -751,6 +788,14 @@ struct PowerFunctor
     explicit PowerFunctor(float power_ = 1.f, float scale_ = 1.f, float shift_ = 0.f)
         : power(power_), scale(scale_), shift(shift_) {}
 
+    bool supportBackend(int backendId, int targetId)
+    {
+        if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
+            return (targetId != DNN_TARGET_OPENCL && targetId != DNN_TARGET_OPENCL_FP16) || power == 1.0;
+        else
+            return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         float a = scale, b = shift, p = power;
@@ -853,6 +898,11 @@ struct ChannelsPReLUFunctor
         scale_umat = scale.getUMat(ACCESS_READ);
     }
 
+    bool supportBackend(int backendId, int)
+    {
+        return backendId == DNN_BACKEND_OPENCV || backendId == DNN_BACKEND_HALIDE;
+    }
+
     void apply(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const
     {
         CV_Assert(scale.isContinuous() && scale.type() == CV_32F);

modules/dnn/src/layers/resize_layer.cpp

@@ -53,8 +53,10 @@ public:
 
     virtual bool supportBackend(int backendId) CV_OVERRIDE
     {
-        return backendId == DNN_BACKEND_OPENCV ||
-               backendId == DNN_BACKEND_INFERENCE_ENGINE && haveInfEngine() && interpolation == "nearest";
+        if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
+            return interpolation == "nearest" && preferableTarget != DNN_TARGET_MYRIAD;
+        else
+            return backendId == DNN_BACKEND_OPENCV;
     }
 
     virtual void finalize(const std::vector<Mat*>& inputs, std::vector<Mat> &outputs) CV_OVERRIDE

modules/dnn/src/layers/slice_layer.cpp

@@ -41,6 +41,7 @@
 //M*/
 
 #include "../precomp.hpp"
+#include "../op_inf_engine.hpp"
 #include "layers_common.hpp"
 #include <opencv2/dnn/shape_utils.hpp>
 
@@ -107,6 +108,12 @@ public:
         }
     }
 
+    virtual bool supportBackend(int backendId) CV_OVERRIDE
+    {
+        return backendId == DNN_BACKEND_OPENCV ||
+               backendId == DNN_BACKEND_INFERENCE_ENGINE && sliceRanges.size() == 1;
+    }
+
     bool getMemoryShapes(const std::vector<MatShape> &inputs,
                             const int requiredOutputs,
                             std::vector<MatShape> &outputs,
@@ -247,6 +254,29 @@ public:
             inpMat(sliceRanges[i]).copyTo(outputs[i]);
         }
     }
+
+    virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >& inputs) CV_OVERRIDE
+    {
+#ifdef HAVE_INF_ENGINE
+        InferenceEngine::DataPtr input = infEngineDataNode(inputs[0]);
+        InferenceEngine::LayerParams lp;
+        lp.name = name;
+        lp.type = "Crop";
+        lp.precision = InferenceEngine::Precision::FP32;
+        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)
+        {
+            ieLayer->axis.push_back(i);
+            ieLayer->offset.push_back(sliceRanges[0][i].start);
+            ieLayer->dim.push_back(sliceRanges[0][i].end - sliceRanges[0][i].start);
+        }
+        return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
+
+#endif  // HAVE_INF_ENGINE
+        return Ptr<BackendNode>();
+    }
 };
 
 Ptr<SliceLayer> SliceLayer::create(const LayerParams& params)

modules/dnn/test/test_backends.cpp

@@ -256,7 +256,8 @@ TEST_P(DNNTestNetwork, OpenPose_pose_mpi_faster_4_stages)
 TEST_P(DNNTestNetwork, OpenFace)
 {
     if (backend == DNN_BACKEND_HALIDE ||
-        backend == DNN_BACKEND_INFERENCE_ENGINE && target != DNN_TARGET_CPU)
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_OPENCL_FP16) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
         throw SkipTestException("");
     processNet("dnn/openface_nn4.small2.v1.t7", "", Size(96, 96), "");
 }
@@ -296,6 +297,20 @@ TEST_P(DNNTestNetwork, DenseNet_121)
     processNet("dnn/DenseNet_121.caffemodel", "dnn/DenseNet_121.prototxt", Size(224, 224), "", "caffe");
 }
 
+TEST_P(DNNTestNetwork, FastNeuralStyle_eccv16)
+{
+    if (backend == DNN_BACKEND_HALIDE ||
+        (backend == DNN_BACKEND_OPENCV && target == DNN_TARGET_OPENCL_FP16) ||
+        (backend == DNN_BACKEND_INFERENCE_ENGINE && target == DNN_TARGET_MYRIAD))
+        throw SkipTestException("");
+    Mat img = imread(findDataFile("dnn/googlenet_1.png", false));
+    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) ? 0.3 : 4e-5;
+    float lInf = (target == DNN_TARGET_OPENCL_FP16 || target == DNN_TARGET_MYRIAD) ? 7.0 : 2e-3;
+    processNet("dnn/fast_neural_style_eccv16_starry_night.t7", "", inp, "", "", l1, lInf);
+}
+
 const tuple<DNNBackend, DNNTarget> testCases[] = {
 #ifdef HAVE_HALIDE
     tuple<DNNBackend, DNNTarget>(DNN_BACKEND_HALIDE, DNN_TARGET_CPU),

modules/dnn/test/test_darknet_importer.cpp

@@ -135,8 +135,6 @@ TEST_P(Test_Darknet_nets, YoloVoc)
 {
     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(1, "detection_out");
 
     std::vector<int> classIds(3);

modules/dnn/test/test_torch_importer.cpp

@@ -296,6 +296,7 @@ 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.