Merge pull request #15032 from l-bat:reduce_mean

* Added support for the ONNX "ReduceMean" Layer. (as this is the same as the GlobalAveragePool)

* Add ReduceMean test

* Fix ONNX importer

* Fix ReduceMean

* Add assert

* Split test

* Fix split test

modules/dnn/src/dnn.cpp

@@ -2170,7 +2170,10 @@ struct Net::Impl
                 if (isAsync)
                     CV_Error(Error::StsNotImplemented, "Default implementation fallbacks in asynchronous mode");
 
-                CV_Assert(layer->supportBackend(DNN_BACKEND_OPENCV));
+                if (!layer->supportBackend(DNN_BACKEND_OPENCV))
+                    CV_Error(Error::StsNotImplemented, format("Layer \"%s\" of type \"%s\" unsupported on OpenCV backend",
+                                                       ld.name.c_str(), ld.type.c_str()));
+
                 if (preferableBackend == DNN_BACKEND_OPENCV && IS_DNN_OPENCL_TARGET(preferableTarget))
                 {
                     std::vector<UMat> umat_inputBlobs = OpenCLBackendWrapper::getUMatVector(ld.inputBlobsWrappers);

modules/dnn/src/layers/layers_common.cpp

@@ -148,13 +148,12 @@ void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kern
                             std::vector<size_t>& pads_begin, std::vector<size_t>& pads_end,
                             std::vector<size_t>& strides, cv::String &padMode)
 {
-    util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode);
-
     globalPooling = params.has("global_pooling") &&
                     params.get<bool>("global_pooling");
 
     if (globalPooling)
     {
+        util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode);
         if(params.has("kernel_h") || params.has("kernel_w") || params.has("kernel_size"))
         {
             CV_Error(cv::Error::StsBadArg, "In global_pooling mode, kernel_size (or kernel_h and kernel_w) cannot be specified");
@@ -171,6 +170,7 @@ void getPoolingKernelParams(const LayerParams &params, std::vector<size_t>& kern
     else
     {
         util::getKernelSize(params, kernel);
+        util::getStrideAndPadding(params, pads_begin, pads_end, strides, padMode, kernel.size());
     }
 }
 

modules/dnn/src/onnx/onnx_importer.cpp

@@ -397,11 +397,33 @@ void ONNXImporter::populateNet(Net dstNet)
             layerParams.set("ceil_mode", layerParams.has("pad_mode"));
             layerParams.set("ave_pool_padded_area", framework_name == "pytorch");
         }
-        else if (layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool")
+        else if (layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool" || layer_type == "ReduceMean")
         {
+            CV_Assert(node_proto.input_size() == 1);
             layerParams.type = "Pooling";
-            layerParams.set("pool", layer_type == "GlobalAveragePool" ? "AVE" : "MAX");
-            layerParams.set("global_pooling", true);
+            layerParams.set("pool", layer_type == "GlobalMaxPool"? "MAX" : "AVE");
+            layerParams.set("global_pooling", layer_type == "GlobalAveragePool" || layer_type == "GlobalMaxPool");
+
+            if (layer_type == "ReduceMean")
+            {
+                if (layerParams.get<int>("keepdims") == 0 || !layerParams.has("axes"))
+                    CV_Error(Error::StsNotImplemented, "Unsupported mode of ReduceMean operation.");
+
+                MatShape inpShape = outShapes[node_proto.input(0)];
+                if (inpShape.size() != 4 && inpShape.size() != 5)
+                    CV_Error(Error::StsNotImplemented, "Unsupported input shape of reduce_mean operation.");
+
+                DictValue axes = layerParams.get("axes");
+                CV_Assert(axes.size() <= inpShape.size() - 2);
+                std::vector<int> kernel_size(inpShape.size() - 2, 1);
+                for (int i = 0; i < axes.size(); i++) {
+                    int axis = axes.get<int>(i);
+                    CV_Assert_N(axis >= 2 + i, axis < inpShape.size());
+                    kernel_size[axis - 2] = inpShape[axis];
+                }
+
+                layerParams.set("kernel_size", DictValue::arrayInt(&kernel_size[0], kernel_size.size()));
+            }
         }
         else if (layer_type == "Slice")
         {
@@ -747,11 +769,13 @@ void ONNXImporter::populateNet(Net dstNet)
             if (axes.size() != 1)
                 CV_Error(Error::StsNotImplemented, "Multidimensional unsqueeze");
 
-            int dims[] = {1, -1};
+            MatShape inpShape = outShapes[node_proto.input(0)];
+            int axis = axes.getIntValue(0);
+            CV_Assert(0 <= axis && axis <= inpShape.size());
+            std::vector<int> outShape = inpShape;
+            outShape.insert(outShape.begin() + axis, 1);
             layerParams.type = "Reshape";
-            layerParams.set("axis", axes.getIntValue(0));
-            layerParams.set("num_axes", 1);
-            layerParams.set("dim", DictValue::arrayInt(&dims[0], 2));
+            layerParams.set("dim", DictValue::arrayInt(&outShape[0], outShape.size()));
         }
         else if (layer_type == "Reshape")
         {

modules/dnn/test/test_onnx_importer.cpp

@@ -162,6 +162,18 @@ TEST_P(Test_ONNX_layers, Clip)
     testONNXModels("clip", npy);
 }
 
+TEST_P(Test_ONNX_layers, ReduceMean)
+{
+    testONNXModels("reduce_mean");
+}
+
+TEST_P(Test_ONNX_layers, ReduceMean3D)
+{
+    if (target != DNN_TARGET_CPU)
+        throw SkipTestException("Only CPU is supported");
+    testONNXModels("reduce_mean3d");
+}
+
 TEST_P(Test_ONNX_layers, MaxPooling_Sigmoid)
 {
     testONNXModels("maxpooling_sigmoid");

modules/dnn/test/test_tf_importer.cpp

@@ -350,11 +350,6 @@ TEST_P(Test_TensorFlow_layers, l2_normalize_3d)
     runTensorFlowNet("l2_normalize_3d");
 }
 
-TEST_P(Test_TensorFlow_layers, Split)
-{
-    runTensorFlowNet("split");
-}
-
 class Test_TensorFlow_nets : public DNNTestLayer {};
 
 TEST_P(Test_TensorFlow_nets, MobileNet_SSD)
@@ -682,6 +677,9 @@ TEST_P(Test_TensorFlow_layers, lstm)
 
 TEST_P(Test_TensorFlow_layers, split)
 {
+    if (backend == DNN_BACKEND_INFERENCE_ENGINE)
+        applyTestTag(CV_TEST_TAG_DNN_SKIP_IE_MYRIAD_2);
+    runTensorFlowNet("split");
     if (backend == DNN_BACKEND_INFERENCE_ENGINE)
         applyTestTag(CV_TEST_TAG_DNN_SKIP_IE);
     runTensorFlowNet("split_equals");