Fixing LRN op to match ONNX 1.2 spec. Size attribute has semantics of diameter and not radius, as was the case before.

Source/CNTKv2LibraryDll/proto/onnx/CNTKToONNX.cpp

@@ -2891,7 +2891,7 @@ void CNTKToONNXHelper::CopyAttributes(const FunctionPtr& src, LotusIR::Node* nod
             auto alpha = (float)src->Attributes()[L"alpha"].Value<double>();
             auto beta = (float)src->Attributes()[L"beta"].Value<double>();
 
-            node->AddAttribute(attributesMap[L"size"], depthRadius);
+            node->AddAttribute(attributesMap[L"size"], 2*depthRadius + 1);
             node->AddAttribute(attributesMap[L"bias"], bias);
             node->AddAttribute(attributesMap[L"alpha"], alpha);
             node->AddAttribute(attributesMap[L"beta"], beta);

Source/CNTKv2LibraryDll/proto/onnx/ONNXToCNTK.cpp

@@ -2019,14 +2019,12 @@ FunctionPtr ONNXToCNTKHelper::CreateFunction(const Node *node, const std::vector
     }
     else if (onnxOpName == "LRN")
     {
-        // TODO: this is experimental code to load Facebook Caffe models.
-        // Operators are added so hopefully there is not further work needed.
-        size_t depthRadius = GetNamedAttributeAsInt64(node, "size");
-        double bias = GetNamedAttributeAsFloat(node, "bias");
-        double alpha = GetNamedAttributeAsFloat(node, "alpha");
-        double beta = GetNamedAttributeAsFloat(node, "beta");
-        FunctionPtr cntkFunction = LocalResponseNormalization(inputs[0],
-                                                              depthRadius, bias, alpha, beta, ToWString(node->Name()));
+        size_t depthRadius = (GetNamedAttributeAsInt64(node, "size") - 1)/2;
+        double bias = static_cast<double>(GetNamedAttributeAsFloat(node, "bias", 1.0f));
+        double alpha = static_cast<double>(GetNamedAttributeAsFloat(node, "alpha", 1e-4f));
+        double beta = static_cast<double>(GetNamedAttributeAsFloat(node, "beta", 0.75f));
+        FunctionPtr cntkFunction = LocalResponseNormalization(inputs[0], 
+            depthRadius, bias, alpha, beta, ToWString(node->Name()));
         return cntkFunction;
     }
     else if (onnxOpName == "AveragePool" || onnxOpName == "MaxPool")

bindings/python/cntk/tests/onnx_op_test.py

@@ -625,7 +625,7 @@ def test_LogSoftmax(tmpdir, dtype):
 @pytest.mark.parametrize("dtype", DType_Config)
 def test_LRN(tmpdir, dtype, device_id):
     if device_id == -1 and dtype == np.float16:
-        pytest.skip('Test is skipped on CPU with float16 data')
+        pytest.skip('Test is skipped on CPU with float16 data, because it uses convolution.')
     device = cntk_device(device_id)
     with C.default_options(dtype=dtype):
         img_shape = (64, 32, 32)