MVN layer using Intel's Inference Engine backend

modules/dnn/include/opencv2/dnn/all_layers.hpp

@@ -489,7 +489,7 @@ CV__DNN_EXPERIMENTAL_NS_BEGIN
         static Ptr<EltwiseLayer> create(const LayerParams &params);
     };
 
-    class CV_EXPORTS BatchNormLayer : public Layer
+    class CV_EXPORTS BatchNormLayer : public ActivationLayer
     {
     public:
         bool hasWeights, hasBias;

modules/dnn/src/dnn.cpp

@@ -1471,6 +1471,8 @@ struct Net::Impl
             {
                 node = layer->initInfEngine(ld.inputBlobsWrappers);
             }
+            else if (node.empty())
+                continue;
 
             CV_Assert(!node.empty());
             ld.backendNodes[preferableBackend] = node;
@@ -1715,40 +1717,41 @@ struct Net::Impl
                 if (preferableBackend != DNN_BACKEND_OPENCV)
                     continue;  // Go to the next layer.
 
-                // For now, OpenCL target support fusion with activation of ReLU/ChannelsPReLU/Power/Tanh
+                while (nextData)
-                if ( !IS_DNN_OPENCL_TARGET(preferableTarget) ||
-                     (IS_DNN_OPENCL_TARGET(preferableTarget) &&
-                         nextData &&
-                        ((nextData->type == "ReLU") ||
-                         (nextData->type == "ChannelsPReLU") ||
-                         (nextData->type == "ReLU6") ||
-                         (nextData->type == "TanH") ||
-                         (nextData->type == "Power"))) )
                 {
+                    // For now, OpenCL target support fusion with activation of ReLU/ChannelsPReLU/Power/Tanh
+                    if (IS_DNN_OPENCL_TARGET(preferableTarget) &&
+                        nextData->type != "ReLU" &&
+                        nextData->type != "ChannelsPReLU" &&
+                        nextData->type != "ReLU6" &&
+                        nextData->type != "TanH" &&
+                        nextData->type != "Power")
+                        break;
 
-                    Ptr<ActivationLayer> nextActivLayer;
+                    Ptr<ActivationLayer> nextActivLayer = nextData->layerInstance.dynamicCast<ActivationLayer>();
-
+                    if (nextActivLayer.empty())
-                    if( nextData )
+                        break;
-                        nextActivLayer = nextData->layerInstance.dynamicCast<ActivationLayer>();
 
-                    if( !nextActivLayer.empty() && pinsToKeep.count(lpNext) == 0
+                    if (currLayer->setActivation(nextActivLayer))
-                            && currLayer->setActivation(nextActivLayer) )
                     {
-                        LayerData *activData = nextData;
                         printf_(("\tfused with %s\n", nextActivLayer->name.c_str()));
-                        activData->skip = true;
+                        nextData->skip = true;
                         ld.outputBlobs = layers[lpNext.lid].outputBlobs;
                         ld.outputBlobsWrappers = layers[lpNext.lid].outputBlobsWrappers;
-
+                        if (nextData->consumers.size() == 1)
-                        if ( IS_DNN_OPENCL_TARGET(preferableTarget) )
-                        {
-                            if ( !activData->consumers.empty() )
                         {
-                                nextData = &layers[activData->consumers[0].lid];
+                            int nextLayerId = nextData->consumers[0].lid;
-                                lpNext = LayerPin(activData->consumers[0].lid, 0);
+                            nextData = &layers[nextLayerId];
+                            lpNext = LayerPin(nextLayerId, 0);
                         }
+                        else
+                        {
+                            nextData = 0;
+                            break;
                         }
                     }
+                    else
+                        break;
                 }
 
                 // fuse convolution layer followed by eltwise + relu

modules/dnn/src/layers/batch_norm_layer.cpp

@@ -268,6 +268,36 @@ public:
         }
     }
 
+    void forwardSlice(const float* srcptr, float* dstptr, int len, size_t planeSize, int cn0, int cn1) const CV_OVERRIDE
+    {
+        for( int cn = cn0; cn < cn1; cn++, srcptr += planeSize, dstptr += planeSize )
+        {
+            int i = 0;
+            float w = weights_.at<float>(cn);
+            float b = bias_.at<float>(cn);
+#if CV_SIMD128
+            v_float32x4 wV = v_setall_f32(w), bV = v_setall_f32(b);
+            for( ; i <= len - 16; i += 16 )
+            {
+                v_float32x4 x0 = v_load(srcptr + i);
+                v_float32x4 x1 = v_load(srcptr + i + 4);
+                v_float32x4 x2 = v_load(srcptr + i + 8);
+                v_float32x4 x3 = v_load(srcptr + i + 12);
+                x0 = v_muladd(x0, w, b);
+                x1 = v_muladd(x1, w, b);
+                x2 = v_muladd(x2, w, b);
+                x3 = v_muladd(x3, w, b);
+                v_store(dstptr + i, x0);
+                v_store(dstptr + i + 4, x1);
+                v_store(dstptr + i + 8, x2);
+                v_store(dstptr + i + 12, x3);
+            }
+#endif
+            for( ; i < len; i++ )
+                dstptr[i] = w * srcptr[i] + b;
+        }
+    }
+
     virtual Ptr<BackendNode> tryAttach(const Ptr<BackendNode>& node) CV_OVERRIDE
     {
         switch (node->backendId)

modules/dnn/src/layers/convolution_layer.cpp

@@ -296,6 +296,9 @@ public:
 
     bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
     {
+        if (!activ.empty() && !layer.empty())
+            return false;
+
         activ = layer;
         if (activ.empty())
             reluslope.clear();

modules/dnn/src/layers/eltwise_layer.cpp

@@ -451,10 +451,15 @@ public:
     }
 
     bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
+    {
+        if (activ.empty() || layer.empty())
         {
             activ = layer;
             return !activ.empty();
         }
+        else
+            return false;
+    }
 
     Ptr<ActivationLayer> activ;
 };

modules/dnn/src/layers/fully_connected_layer.cpp

@@ -134,10 +134,15 @@ public:
     }
 
     virtual bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
+    {
+        if (activ.empty() || layer.empty())
         {
             activ = layer;
             return !activ.empty();
         }
+        else
+            return false;
+    }
 
     class FullyConnected : public ParallelLoopBody
     {

modules/dnn/src/layers/mvn_layer.cpp

@@ -42,6 +42,7 @@
 
 #include "../precomp.hpp"
 #include "layers_common.hpp"
+#include "../op_inf_engine.hpp"
 #include <opencv2/dnn/shape_utils.hpp>
 
 #ifdef HAVE_OPENCL
@@ -66,27 +67,25 @@ public:
         fuse_batch_norm = false;
         fuse_relu = false;
         relu_slope = 0.f;
+        zeroDev = false;
     }
 
     Mat scale, shift;
     bool fuse_batch_norm;
 
-    virtual bool tryFuse(Ptr<Layer>& top) CV_OVERRIDE
+    Ptr<ReLULayer> activ_relu;
+    float relu_slope;
+    bool fuse_relu;
+    bool zeroDev;  // TODO: Doesn't considered in Intel's Inference Engine backend.
+    bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
     {
-        if (!fuse_batch_norm)
+        if (!layer.empty() && !fuse_relu && !fuse_batch_norm)
         {
-            top->getScaleShift(scale, shift);
+            layer->getScaleShift(scale, shift);
             fuse_batch_norm = !scale.empty() || !shift.empty();
             return fuse_batch_norm;
         }
-        return false;
-    }
 
-    Ptr<ReLULayer> activ_relu;
-    float relu_slope;
-    bool fuse_relu;
-    bool setActivation(const Ptr<ActivationLayer>& layer) CV_OVERRIDE
-    {
         if (!layer.empty() && preferableTarget == DNN_TARGET_OPENCL)
         {
             activ_relu = layer.dynamicCast<ReLULayer>();
@@ -97,6 +96,23 @@ public:
         return fuse_relu;
     }
 
+    void finalize(const std::vector<Mat*> &inputs, std::vector<Mat> &outputs) CV_OVERRIDE
+    {
+        int splitDim = (acrossChannels) ? 1 : 2;
+        int i, newRows = 1;
+        for( i = 0; i < splitDim; i++ )
+            newRows *= inputs[0]->size[i];
+        zeroDev = inputs[0]->total() == newRows;
+    }
+
+    virtual bool supportBackend(int backendId) CV_OVERRIDE
+    {
+        if (backendId == DNN_BACKEND_INFERENCE_ENGINE)
+            return !zeroDev && (preferableTarget == DNN_TARGET_CPU || eps <= 1e-7f);
+        else
+            return backendId == DNN_BACKEND_OPENCV;
+    }
+
 #ifdef HAVE_OPENCL
     bool fast_forward_ocl(std::vector<UMat> &inputs, std::vector<UMat> &outputs)
     {
@@ -324,6 +340,22 @@ public:
         }
     }
 
+    virtual Ptr<BackendNode> initInfEngine(const std::vector<Ptr<BackendWrapper> >&) CV_OVERRIDE
+    {
+#ifdef HAVE_INF_ENGINE
+        InferenceEngine::LayerParams lp;
+        lp.name = name;
+        lp.type = "MVN";
+        lp.precision = InferenceEngine::Precision::FP32;
+        std::shared_ptr<InferenceEngine::MVNLayer> ieLayer(new InferenceEngine::MVNLayer(lp));
+        ieLayer->params["across_channels"] = acrossChannels ? "1" : "0";
+        ieLayer->params["normalize_variance"] = normVariance ? "1" : "0";
+        ieLayer->params["eps"] = format("%f", eps);
+        return Ptr<BackendNode>(new InfEngineBackendNode(ieLayer));
+#endif  // HAVE_INF_ENGINE
+        return Ptr<BackendNode>();
+    }
+
     virtual int64 getFLOPS(const std::vector<MatShape> &inputs,
                            const std::vector<MatShape> &outputs) const CV_OVERRIDE
     {

modules/dnn/test/test_tf_importer.cpp

@@ -165,12 +165,6 @@ TEST_P(Test_TensorFlow_layers, batch_norm)
     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");
 }