From 5170d4a2cea2a5618bd87b7c7e66115ef48f32b1 Mon Sep 17 00:00:00 2001
From: Zhou Wang <zhouwang@microsoft.com>
Date: Thu, 26 Jan 2017 21:25:33 +0100
Subject: [PATCH 1/7] make CNTKLibraryCSEvalExampletest project optional for
 build

use the same conditional compilation pattern as in C++
---
 .../CNTKLibraryCSEvalExamplesTest.csproj              | 10 +++++++---
 .../CNTKLibraryManagedDll.csproj                      | 11 ++++++-----
 bindings/csharp/Swig/CNTKLibraryCSBinding.vcxproj     | 11 ++++++-----
 3 files changed, 19 insertions(+), 13 deletions(-)

diff --git a/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj b/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj
index 27357a01c..35ac9485a 100644
--- a/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj
+++ b/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj
@@ -1,5 +1,5 @@
 <?xml version="1.0" encoding="utf-8"?>
-<Project ToolsVersion="14.0" DefaultTargets="Build" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+<Project ToolsVersion="14.0" DefaultTargets="Build" InitialTargets="CheckDependencies" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <Import Project="$(MSBuildExtensionsPath)\$(MSBuildToolsVersion)\Microsoft.Common.props" Condition="Exists('$(MSBuildExtensionsPath)\$(MSBuildToolsVersion)\Microsoft.Common.props')" />
   <PropertyGroup>
     <Configuration Condition=" '$(Configuration)' == '' ">Debug</Configuration>
@@ -63,7 +63,7 @@
     <Reference Include="System" />
     <Reference Include="System.Drawing" />
   </ItemGroup>
-  <ItemGroup Condition="Exists('$(SWIG_PATH)\swig.exe')">
+  <ItemGroup>
     <Compile Include="Properties\AssemblyInfo.cs" />
   </ItemGroup>
   <ItemGroup>
@@ -84,7 +84,11 @@
     </Compile>
     <Compile Include="Program.cs" />
   </ItemGroup>
-  <Import Project="$(MSBuildToolsPath)\Microsoft.CSharp.targets" Condition="Exists('$(SWIG_PATH)\swig.exe')" />
+  <Import Project="$(MSBuildToolsPath)\Microsoft.CSharp.targets" />
+  <Target Name="Build" Condition="Exists('$(SWIG_PATH)\swig.exe')" DependsOnTargets="$(BuildDependsOn)" />
+  <Target Name="CheckDependencies">
+    <Warning Condition="!Exists('$(SWIG_PATH)\swig.exe')" Text="The project requires SWIG to be installed. Please see https://github.com/Microsoft/CNTK/wiki/Setup-CNTK-on-Windows#optional-cntk-cnet-library-for-evaluation for installation instructions." />
+  </Target>
   <!-- To modify your build process, add your task inside one of the targets below and uncomment it. 
        Other similar extension points exist, see Microsoft.Common.targets.
   <Target Name="BeforeBuild">
diff --git a/bindings/csharp/CNTKLibraryManagedDll/CNTKLibraryManagedDll.csproj b/bindings/csharp/CNTKLibraryManagedDll/CNTKLibraryManagedDll.csproj
index 1d9cf7a54..1a183f79d 100644
--- a/bindings/csharp/CNTKLibraryManagedDll/CNTKLibraryManagedDll.csproj
+++ b/bindings/csharp/CNTKLibraryManagedDll/CNTKLibraryManagedDll.csproj
@@ -1,5 +1,5 @@
 <?xml version="1.0" encoding="utf-8"?>
-<Project ToolsVersion="14.0" DefaultTargets="Build" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+<Project ToolsVersion="14.0" DefaultTargets="Build" InitialTargets="CheckDependencies" xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
   <Import Project="$(MSBuildExtensionsPath)\$(MSBuildToolsVersion)\Microsoft.Common.props" Condition="Exists('$(MSBuildExtensionsPath)\$(MSBuildToolsVersion)\Microsoft.Common.props')" />
   <PropertyGroup>
     <Configuration Condition=" '$(Configuration)' == '' ">Debug</Configuration>
@@ -63,7 +63,7 @@
     <Reference Include="System" />
     <Reference Include="System.Drawing" />
   </ItemGroup>
-  <ItemGroup Condition="Exists('$(SWIG_PATH)\swig.exe')">
+  <ItemGroup>
     <Compile Include="Properties\AssemblyInfo.cs" />
   </ItemGroup>
   <ItemGroup>
@@ -98,10 +98,11 @@
     <Compile Include="SwigProxyClasses\VariableKind.cs" />
     <Compile Include="SwigProxyClasses\VariableVector.cs" />
   </ItemGroup>
-  <Target Name="Build" Condition="!Exists('$(SWIG_PATH)\swig.exe')">
-    <Warning Condition="!Exists('$(SWIG_PATH)\swig.exe')" Text="The project requires SWIG to be installed." />
+  <Import Project="$(MSBuildToolsPath)\Microsoft.CSharp.targets" />
+  <Target Name="Build" Condition="Exists('$(SWIG_PATH)\swig.exe')" DependsOnTargets="$(BuildDependsOn)" />
+  <Target Name="CheckDependencies">
+    <Warning Condition="!Exists('$(SWIG_PATH)\swig.exe')" Text="The project requires SWIG to be installed. Please see https://github.com/Microsoft/CNTK/wiki/Setup-CNTK-on-Windows#optional-cntk-cnet-library-for-evaluation for installation instructions." />
   </Target>
-  <Import Project="$(MSBuildToolsPath)\Microsoft.CSharp.targets" Condition="Exists('$(SWIG_PATH)\swig.exe')" />
   <!-- To modify your build process, add your task inside one of the targets below and uncomment it. 
        Other similar extension points exist, see Microsoft.Common.targets.
   <Target Name="BeforeBuild">
diff --git a/bindings/csharp/Swig/CNTKLibraryCSBinding.vcxproj b/bindings/csharp/Swig/CNTKLibraryCSBinding.vcxproj
index 2dcd780ff..322661f04 100644
--- a/bindings/csharp/Swig/CNTKLibraryCSBinding.vcxproj
+++ b/bindings/csharp/Swig/CNTKLibraryCSBinding.vcxproj
@@ -101,7 +101,7 @@
   </ItemDefinitionGroup>
   <PropertyGroup Label="UserMacros" />
   <PropertyGroup />
-  <ItemGroup Condition="$(HasSwig)">
+  <ItemGroup>
     <CustomBuild Include="cntk_cs.i">
       <FileType>Document</FileType>
       <Command>$(SWIG_PATH)\swig.exe -c++ -csharp -DMSC_VER -I$(SolutionDir)Source\CNTKv2LibraryDll\API -I$(SolutionDir)bindings\common -namespace CNTK -outdir  $(SolutionDir)bindings\csharp\CNTKLibraryManagedDll\SwigProxyClasses -dllimport CNTKLibraryCSBinding cntk_cs.i</Command>
@@ -109,10 +109,10 @@
       <AdditionalInputs>$(SolutionDir)Source\CNTKv2LibraryDll\API\CNTKLibrary.h;$(SolutionDir)bindings\common\CNTKValueExtend.i;std_unordered_map.i</AdditionalInputs>
     </CustomBuild>
   </ItemGroup>
-  <ItemGroup Condition="$(HasSwig)">
+  <ItemGroup>
     <ClCompile Include="cntk_cs_wrap.cxx" />
   </ItemGroup>
-  <ItemGroup Condition="$(HasSwig)">
+  <ItemGroup>
     <ClInclude Include="..\..\..\Source\CNTKv2LibraryDll\API\CNTKLibrary.h" />
     <ClInclude Include="..\..\..\Source\CNTKv2LibraryDll\API\CNTKLibraryInternals.h" />
     <ClInclude Include="cntk_cs_wrap.h" />
@@ -120,9 +120,10 @@
   <ItemGroup>
     <None Include="..\..\common\CNTKValueExtend.i" />
   </ItemGroup>
-  <Target Name="CheckDependencies" Condition="!$(HasSwig)">
-    <Warning Condition="!$(HasSwig)" Text="The project requires SWIG to be installed." />
+  <Target Name="CheckDependencies">
+    <Warning Condition="!$(HasSwig)" Text="The project requires SWIG to be installed. Please see https://github.com/Microsoft/CNTK/wiki/Setup-CNTK-on-Windows#optional-cntk-cnet-library-for-evaluation for installation instructions." />
   </Target>
   <Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
+  <Target Name="Build" Condition="$(HasSwig)" Outputs="$(TargetPath)" DependsOnTargets="$(BuildDependsOn)" />
   <ImportGroup Label="ExtensionTargets" />
 </Project>
\ No newline at end of file

From e0feaefcf0a2c742391f82cc204d27a9aa123e76 Mon Sep 17 00:00:00 2001
From: Amit Agarwal <amitaga@microsoft.com>
Date: Wed, 1 Feb 2017 09:07:57 -0800
Subject: [PATCH 2/7] CNTK v2 library: 1. Allow parameter and constant objects
 to directly use a preallocated NDArrayView instead of always creating a clone
 2. Fix an issue with loading of legacy convolution models where the parameter
 tensor is shaped as 2D instead of the real convolution map shape

---
 Source/CNTKv2LibraryDll/API/CNTKLibrary.h      | 11 ++++++++---
 Source/CNTKv2LibraryDll/BackCompat.cpp         |  2 +-
 Tests/UnitTests/V2LibraryTests/TensorTests.cpp |  4 ++--
 3 files changed, 11 insertions(+), 6 deletions(-)

diff --git a/Source/CNTKv2LibraryDll/API/CNTKLibrary.h b/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
index 6f2e51364..d7561789b 100644
--- a/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
+++ b/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
@@ -1854,8 +1854,11 @@ private:
 
     private:
         explicit Parameter(const NDArrayViewPtr& value, const std::wstring& name, const std::wstring& uid)
-            : Variable(value->Shape(), VariableKind::Parameter, value->GetDataType(), value->DeepClone(false), true, {}, name, uid)
-        {}
+            : Variable(value->Shape(), VariableKind::Parameter, value->GetDataType(), value, true, {}, name, uid)
+        {
+            if (value->IsReadOnly())
+                InvalidArgument("Parameter cannot be constructed from a read-only NDArrayView value; you can create a non read-only clone of the value and use that instead!");
+        }
     };
 
     // Implementation note: The Variable type is a value type and not polymorphic in nature. 
@@ -1936,7 +1939,7 @@ private:
 
     private:
         Constant(const NDArrayViewPtr& value, const std::wstring& name, const std::wstring& uid)
-            : Variable(value->Shape(), VariableKind::Constant, value->GetDataType(), value->DeepClone(), false, {}, name, uid)
+            : Variable(value->Shape(), VariableKind::Constant, value->GetDataType(), value, false, {}, name, uid)
         {}
 
         ///
@@ -2596,6 +2599,7 @@ namespace CNTK
         CNTK_API virtual void Backward(const BackPropStatePtr& state,
                                        const std::unordered_map<Variable, ValuePtr>& rootGradientValues,
                                        std::unordered_map<Variable, ValuePtr>& backPropagatedGradientValuesForInputs);
+
         ///
         /// Returns the name of the operation that this Function denotes
         ///
@@ -2851,6 +2855,7 @@ namespace CNTK
 
         // Returns a outputs without ref-counting the owner.
         CNTK_API std::vector<Variable>& RawOutputs() const;
+
     private:
         CNTK_API std::shared_ptr<std::vector<std::pair<Variable, Variable>>> BlockArgumentsMappingImpl() const;
 
diff --git a/Source/CNTKv2LibraryDll/BackCompat.cpp b/Source/CNTKv2LibraryDll/BackCompat.cpp
index f2ce2ce4b..8045e5cf4 100644
--- a/Source/CNTKv2LibraryDll/BackCompat.cpp
+++ b/Source/CNTKv2LibraryDll/BackCompat.cpp
@@ -311,7 +311,7 @@ namespace CNTK
                     // tensor dimensions flattended into the column dimension of the 2D paramater matrix
                     // We need to recover the actual tensor shape of the parameter in this case
                     auto& convolutionMapVar = inputVars[0];
-                    if (convolutionNode->IsConvolution2D())
+                    if (convolutionNode->IsConvolution2D() || (convolutionMapVar.Shape().Rank() == 2))
                     {
                         assert(convolutionMapVar.Shape().Rank() == 2);
                         assert(convolutionMapVar.IsConstant() || convolutionMapVar.IsParameter());
diff --git a/Tests/UnitTests/V2LibraryTests/TensorTests.cpp b/Tests/UnitTests/V2LibraryTests/TensorTests.cpp
index 569149e1a..d56b87357 100644
--- a/Tests/UnitTests/V2LibraryTests/TensorTests.cpp
+++ b/Tests/UnitTests/V2LibraryTests/TensorTests.cpp
@@ -41,8 +41,8 @@ void TestTensorPlus(size_t numAxesLeftOperand, size_t numAxesRightOperand, const
     Variable leftInputVar, rightInputVar;
     if (useConstantInputsOnly)
     {
-        leftInputValue = leftInputValue->DeepClone(device, true);
-        rightInputValue = rightInputValue->DeepClone(device, true);
+        leftInputValue = leftInputValue->DeepClone(device, false);
+        rightInputValue = rightInputValue->DeepClone(device, false);
 
         leftInputVar = Parameter(leftInputValue, L"leftInput");
         rightInputVar = Parameter(rightInputValue, L"rightInput");

From cf6b968bed903b38a3ae95615c99f657dbbb59e6 Mon Sep 17 00:00:00 2001
From: Amit Agarwal <amitaga@microsoft.com>
Date: Wed, 1 Feb 2017 13:41:37 -0800
Subject: [PATCH 3/7] CNTK v2 library: Add support for enumerating inputs of
 Functions per the python operand order where samples are row vectors instead
 of column vectors and thus times operation operands are reversed in order

---
 Source/CNTKv2LibraryDll/API/CNTKLibrary.h     |  6 ++---
 Source/CNTKv2LibraryDll/CompositeFunction.h   | 18 +++++++--------
 Source/CNTKv2LibraryDll/Function.cpp          | 20 ++++++++++++----
 bindings/python/cntk/ops/functions.py         |  6 +----
 .../python/cntk/ops/tests/function_tests.py   | 23 +++++++++++++++++++
 5 files changed, 52 insertions(+), 21 deletions(-)

diff --git a/Source/CNTKv2LibraryDll/API/CNTKLibrary.h b/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
index d7561789b..ae693e6bc 100644
--- a/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
+++ b/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
@@ -2734,9 +2734,9 @@ namespace CNTK
         ///
         /// Returns all Input variables of 'this' Function.
         ///
-        std::vector<Variable> Inputs() const
+        std::vector<Variable> Inputs(bool pythonOperandOrder = false) const
         {
-            return *(InputsImpl().get());
+            return *(InputsImpl(pythonOperandOrder).get());
         }
 
         ///
@@ -2880,7 +2880,7 @@ namespace CNTK
             return filteredInputs;
         }
 
-        CNTK_API std::shared_ptr<std::vector<Variable>> InputsImpl() const;
+        CNTK_API std::shared_ptr<std::vector<Variable>> InputsImpl(bool pythonOperandOrder = false) const;
         CNTK_API std::shared_ptr<std::vector<Variable>> OutputsImpl() const;
 
         void ValidateOrUpdateOutputs(std::unordered_map<const Function*, size_t>& visitedFunctions, bool& recurrentNodeOutputModified);
diff --git a/Source/CNTKv2LibraryDll/CompositeFunction.h b/Source/CNTKv2LibraryDll/CompositeFunction.h
index 2dc7deb24..74a29facf 100644
--- a/Source/CNTKv2LibraryDll/CompositeFunction.h
+++ b/Source/CNTKv2LibraryDll/CompositeFunction.h
@@ -146,29 +146,29 @@ namespace CNTK
         }
 
         template <typename FunctionType>
-        static void PreorderTraverseVariables(const FunctionPtr& rootFunction, const FunctionType& functor)
+        static void PreorderTraverseVariables(const FunctionPtr& rootFunction, const FunctionType& functor, bool pythonOperandOrder = false)
         {
             std::unordered_set<FunctionPtr> visitedFunctions;
-            PreorderTraverseVariables(rootFunction, visitedFunctions, functor);
+            PreorderTraverseVariables(rootFunction, visitedFunctions, functor, pythonOperandOrder);
         }
 
         // Recursively traverses the Function graph underlying the 'rootFunction' invoking the provided functor for all visited nodes in the graph.
         template <typename FunctionType>
-        static void PreorderTraverseVariables(const FunctionPtr& rootFunction, std::unordered_set<FunctionPtr>& visitedFunctions, const FunctionType& functor)
+        static void PreorderTraverseVariables(const FunctionPtr& rootFunction, std::unordered_set<FunctionPtr>& visitedFunctions, const FunctionType& functor, bool pythonOperandOrder = false)
         {
             visitedFunctions.insert(rootFunction);
             auto rootFunctionOutputs = rootFunction->InitOutputs();
             for (const auto& rootOutput : rootFunctionOutputs)
                 functor(rootOutput);
 
-            auto rootFunctionInputs = rootFunction->Inputs();
+            auto rootFunctionInputs = rootFunction->Inputs(pythonOperandOrder);
             for (const auto& rootInput : rootFunctionInputs)
             {
                 functor(rootInput);
                 if (rootInput.IsOutput() && visitedFunctions.find(rootInput.Owner()) == visitedFunctions.end())
                 {
                     const auto& function = rootInput.Owner();
-                    PreorderTraverseVariables(function, visitedFunctions, functor);
+                    PreorderTraverseVariables(function, visitedFunctions, functor, pythonOperandOrder);
                 }
             }
         }
@@ -201,11 +201,11 @@ namespace CNTK
             m_allPrimitiveFunctions(std::move(allPrimitiveFunctions)), m_networkMatricesAllocated(false)
         {}
 
-        std::vector<Variable> DetermineInputs() const
+        std::vector<Variable> DetermineInputs(bool pythonOperandOrder = false) const
         {
             const auto& root = RootFunction();
             std::unordered_set<FunctionPtr> visitedFunctions;
-            return DetermineInputs(root, visitedFunctions);
+            return DetermineInputs(root, visitedFunctions, pythonOperandOrder);
         }
 
          // Recursively traverses the Function graph and populates the provided set of functions.
@@ -216,7 +216,7 @@ namespace CNTK
         }
 
         // Recursively traverses the Function graph underlying the 'rootFunction' to determine all the leaves (aka inputs) of the graph
-        static std::vector<Variable> DetermineInputs(const FunctionPtr& rootFunction, std::unordered_set<FunctionPtr>& visitedFunctions)
+        static std::vector<Variable> DetermineInputs(const FunctionPtr& rootFunction, std::unordered_set<FunctionPtr>& visitedFunctions, bool pythonOperandOrder = false)
         {
             vector<FunctionPtr> functions;
             std::vector<Variable> inputs;
@@ -227,7 +227,7 @@ namespace CNTK
                     inputs.push_back(var);
                     uniqueInputs.insert(var);
                 }
-           });
+           }, pythonOperandOrder);
 
             return inputs;
         }
diff --git a/Source/CNTKv2LibraryDll/Function.cpp b/Source/CNTKv2LibraryDll/Function.cpp
index a3943512b..226e6c3e7 100644
--- a/Source/CNTKv2LibraryDll/Function.cpp
+++ b/Source/CNTKv2LibraryDll/Function.cpp
@@ -51,14 +51,26 @@ namespace CNTK
         return const_cast<Function*>(this)->InitOutputs();
     }
 
-    std::shared_ptr<std::vector<Variable>> Function::InputsImpl() const
+    std::shared_ptr<std::vector<Variable>> Function::InputsImpl(bool pythonOperandOrder) const
     {
-        const CompositeFunction* compositeFunction = dynamic_cast<const CompositeFunction*>(this);
         std::vector<Variable> inputs;
+
+        const CompositeFunction* compositeFunction = dynamic_cast<const CompositeFunction*>(this);
         if (compositeFunction == nullptr)
-            inputs = m_inputs;
+        {
+            // For the Times and TransposeTimes primitive functions, if we want the python operand order
+            // then we need to reorder the operands as stored in m_inputs
+            const PrimitiveFunction* primitiveFunction = dynamic_cast<const PrimitiveFunction*>(this);
+            if (pythonOperandOrder && primitiveFunction && ((primitiveFunction->OpType() == PrimitiveOpType::Times) || (primitiveFunction->OpType() == PrimitiveOpType::TransposeTimes)))
+            {
+                assert(m_inputs.size() == 2);
+                inputs = { m_inputs[1], m_inputs[0] };
+            }
+            else
+                inputs = m_inputs;
+        }
         else
-            inputs = compositeFunction->DetermineInputs();
+            inputs = compositeFunction->DetermineInputs(pythonOperandOrder);
 
         return std::shared_ptr<std::vector<Variable>>(new std::vector<Variable>(std::move(inputs)), [](std::vector<Variable>* ptr) { delete ptr; });
     }
diff --git a/bindings/python/cntk/ops/functions.py b/bindings/python/cntk/ops/functions.py
index 28503dac4..189f8f896 100644
--- a/bindings/python/cntk/ops/functions.py
+++ b/bindings/python/cntk/ops/functions.py
@@ -391,11 +391,7 @@ class Function(cntk_py.Function):
         '''
         List of all input variables of this function.
         '''
-        input_nodes = super(Function, self).inputs()
-        if self.is_primitive and self.root_function.op_name in ['Times', 'TransposeTimes']:
-            input_nodes = tuple(reversed(input_nodes))
-
-        return input_nodes
+        return super(Function, self).inputs(True)
 
     @property
     def name(self):
diff --git a/bindings/python/cntk/ops/tests/function_tests.py b/bindings/python/cntk/ops/tests/function_tests.py
index 29dc87dca..e132670bf 100644
--- a/bindings/python/cntk/ops/tests/function_tests.py
+++ b/bindings/python/cntk/ops/tests/function_tests.py
@@ -229,3 +229,26 @@ def test_as_composite():
     assert(composite.root_function.name == func_name)
     composite = as_composite(t_plus_b)
     assert(composite.root_function.name == func_name)
+
+def test_input_order():
+    input_dim = 1
+    proj_dim = 2
+    x = input_variable((input_dim,), name='x')
+    b = parameter((proj_dim), name='b')
+    w = parameter((input_dim, proj_dim), name='w')
+    func_name = 't_plus_b'
+    t = times(x, w)
+    t_plus_b = plus(t, b, name=func_name)
+
+    def compare_var_names(vars, names): 
+        num_vars = len(vars)
+        for i in range(num_vars):
+            if (vars[i].name != names[i]):
+                return False
+
+        return True
+
+    assert compare_var_names(t.root_function.inputs, ['x', 'w'])
+    assert compare_var_names(t.inputs, ['x', 'w'])
+    assert compare_var_names(t_plus_b.inputs, ['x', 'w', 'b'])
+

From 566355231c29c8bff8444315293e4fc05fa8d3e8 Mon Sep 17 00:00:00 2001
From: Amit Agarwal <amitaga@microsoft.com>
Date: Wed, 1 Feb 2017 21:14:07 -0800
Subject: [PATCH 4/7] CNTK v2 library: Add support for precompute nodes in
 LoadLegacyModel

---
 Makefile                                      |  1 +
 Source/CNTKv2LibraryDll/BackCompat.cpp        | 54 ++++++++++----
 Source/CNTKv2LibraryDll/Function.cpp          |  5 +-
 .../CNTKv2Library/UnitTests/run-test          | 14 ++++
 Tests/UnitTests/V2LibraryTests/Common.h       | 43 +++++++++++
 .../V2LibraryTests/LoadLegacyModelTests.cpp   | 73 +++++++++++++++++++
 Tests/UnitTests/V2LibraryTests/Main.cpp       |  4 +-
 .../TruncatedLSTMAcousticModel.cpp            | 45 +-----------
 .../V2LibraryTests/V2LibraryTests.vcxproj     |  1 +
 .../V2LibraryTests.vcxproj.filters            |  3 +
 10 files changed, 183 insertions(+), 60 deletions(-)
 create mode 100644 Tests/UnitTests/V2LibraryTests/LoadLegacyModelTests.cpp

diff --git a/Makefile b/Makefile
index 38e5b29b3..5af3ef239 100644
--- a/Makefile
+++ b/Makefile
@@ -506,6 +506,7 @@ CNTKLIBRARY_TESTS_SRC =\
 	$(CNTKLIBRARY_TESTS_SRC_PATH)/BlockTests.cpp \
 	$(CNTKLIBRARY_TESTS_SRC_PATH)/TensorTests.cpp \
 	$(CNTKLIBRARY_TESTS_SRC_PATH)/ValueTests.cpp \
+	$(CNTKLIBRARY_TESTS_SRC_PATH)/LoadLegacyModelTests.cpp \
 	$(CNTKLIBRARY_TESTS_SRC_PATH)/TrainerTests.cpp \
 	$(CNTKLIBRARY_TESTS_SRC_PATH)/CifarResNet.cpp \
 	$(CNTKLIBRARY_TESTS_SRC_PATH)/SerializationTests.cpp \
diff --git a/Source/CNTKv2LibraryDll/BackCompat.cpp b/Source/CNTKv2LibraryDll/BackCompat.cpp
index 8045e5cf4..13cd57c21 100644
--- a/Source/CNTKv2LibraryDll/BackCompat.cpp
+++ b/Source/CNTKv2LibraryDll/BackCompat.cpp
@@ -20,6 +20,8 @@
 #include "ReshapingNodes.h"
 #include "DeprecatedNodes.h"
 #include "RNNNodes.h"
+#include "PreComputeNodes.h"
+#include "DeprecatedNodes.h"
 
 using namespace Microsoft::MSR::CNTK;
 
@@ -49,9 +51,7 @@ namespace CNTK
 
                 Variable var;
                 if (node->IsLeaf())
-                {
                     var = ResolveLeaf<ElementType>(node);
-                }
                 else
                 {
                     // This is a non-leaf node and maps to a primitive Function
@@ -78,6 +78,22 @@ namespace CNTK
             }
 
         private:
+
+            template<class ElementType>
+            Variable CreateParameterOrConstantFromNodeValue(const ComputationNodeBasePtr& node, bool isConstant)
+            {
+                auto& matrix = node->As<ComputationNode<ElementType>>()->Value();
+                auto tensorView = new TensorView<ElementType>(std::make_shared<Matrix<ElementType>>(matrix.AsReference()), AsTensorViewShape(node->GetSampleLayout()));
+                NDArrayViewPtr value = MakeSharedObject<NDArrayView>(AsDataType<ElementType>(), AsDeviceDescriptor(matrix.GetDeviceId()), AsStorageFormat(matrix.GetFormat()), AsNDShape(node->GetSampleLayout()), false, tensorView);
+
+                auto kind = isConstant ? VariableKind::Constant : VariableKind::Parameter;
+
+                std::wstring varUid, varName;
+                std::tie(varUid, varName) = UidAndNameFromCNTKInternalNodeName(node->NodeName(), kind);
+
+                return isConstant ? (Variable)Constant(value, varName, varUid) : Parameter(value, varName, varUid);
+            }
+
             template<class ElementType>
             Variable ResolveLeaf(const ComputationNodeBasePtr& node)
             {
@@ -104,13 +120,7 @@ namespace CNTK
                 if (node->Is<LearnableParameter<ElementType>>())
                 {
                     bool isConstant = (node->GetLearningRateMultiplier() == 0);
-                    auto& matrix = node->As<ComputationNode<ElementType>>()->Value();
-                    auto tensorView = new TensorView<ElementType>(std::make_shared<Matrix<ElementType>>(matrix.AsReference()), AsTensorViewShape(node->GetSampleLayout()));
-                    NDArrayViewPtr value = MakeSharedObject<NDArrayView>(AsDataType<ElementType>(), AsDeviceDescriptor(matrix.GetDeviceId()), AsStorageFormat(matrix.GetFormat()), variableShape, false, tensorView);
-
-                    auto kind = isConstant ? VariableKind::Constant : VariableKind::Parameter;
-                    std::tie(varUid, varName) = UidAndNameFromCNTKInternalNodeName(node->NodeName(), kind);
-                    return isConstant ? (Variable)Constant(value, varName, varUid) : Parameter(value, varName, varUid);
+                    return CreateParameterOrConstantFromNodeValue<ElementType>(node, isConstant);
                 }
 
                 LogicError("CNTK::LoadLegacyModel: Unsupported legacy CNTK node named '%S'", node->NodeName().c_str());
@@ -444,6 +454,24 @@ namespace CNTK
 
                     opType = PrimitiveOpType::EditDistanceError;
                 }
+                else if ((node->OperationName() == OperationNameOf(MeanNode)) || (node->OperationName() == OperationNameOf(InvStdDevNode)))
+                {
+                    auto precomputeNode = node->As<MeanInvStdDevNodeBase<ElementType>>();
+                    if (!precomputeNode->HasComputed())
+                        InvalidArgument("Loading a CNTK legacy V1 model containing a Mean/InvStdDev precompute node, whose computation is unfinished, is not supported!");
+
+                    return CreateParameterOrConstantFromNodeValue<ElementType>(node, /* isConstant =*/ true);
+                }
+                else if (node->OperationName() == OperationNameOf(PerDimMeanVarNormalizationNode))
+                {
+                    auto meanValue = Constant(inputVars[1]).Value();
+                    auto invStdDevValue = Constant(inputVars[2]).Value();
+
+                    std::wstring uid, name;
+                    std::tie(uid, name) = UidAndNameFromCNTKInternalNodeName(node->NodeName());
+
+                    return PerDimMeanVarianceNormalize(inputVars[0], meanValue, invStdDevValue, name);
+                }
                 else
                     LogicError("Unsupported ComputationNode with OperationName='%S' found when loading legacy CNTK model", node->OperationName().c_str());
 
@@ -500,16 +528,16 @@ namespace CNTK
 
                 if (ComputationNetwork::IsNodePtr<ComputationNode<float>>(rootNode))
                 {
-                    rootVariables.push_back(resolver.GetVariable<float>(rootNode).Owner());
+                    auto var = resolver.GetVariable<float>(rootNode);
+                    rootVariables.push_back(var.IsOutput() ? (Variable)var.Owner() : var);
                 }
                 else if (ComputationNetwork::IsNodePtr<ComputationNode<double>>(rootNode))
                 {
-                    rootVariables.push_back(resolver.GetVariable<double>(rootNode).Owner());
+                    auto var = resolver.GetVariable<double>(rootNode);
+                    rootVariables.push_back(var.IsOutput() ? (Variable)var.Owner() : var);
                 }
                 else
-                {
                     LogicError("LoadLegacyModel(): invalid computation node element type.");
-                }
             }
 
             auto rootComposite = Combine(rootVariables);
diff --git a/Source/CNTKv2LibraryDll/Function.cpp b/Source/CNTKv2LibraryDll/Function.cpp
index 226e6c3e7..974fe1de9 100644
--- a/Source/CNTKv2LibraryDll/Function.cpp
+++ b/Source/CNTKv2LibraryDll/Function.cpp
@@ -1127,12 +1127,11 @@ namespace CNTK
 
     FunctionPtr PerDimMeanVarianceNormalize(const Variable& operand, const NDArrayViewPtr& mean, const NDArrayViewPtr& invStdDev, const std::wstring& name)
     {
-        // TODO: Should this too be encapsulated as a block?
-
+        auto operandPlaceholder = PlaceholderVariable(L"operand");
         Constant meanVar(mean);
         Constant invStdDevVar(invStdDev);
 
-        return ElementTimes(Minus(operand, meanVar), invStdDevVar, name);
+        return AsBlock(std::move(ElementTimes(Minus(operandPlaceholder, meanVar), invStdDevVar)), { { operandPlaceholder, operand } }, L"PerDimMeanVarianceNormalize", name);
     }
 
     FunctionPtr Convolution(const Variable& convolutionMap,
diff --git a/Tests/EndToEndTests/CNTKv2Library/UnitTests/run-test b/Tests/EndToEndTests/CNTKv2Library/UnitTests/run-test
index 1dfde3aad..c16e23047 100755
--- a/Tests/EndToEndTests/CNTKv2Library/UnitTests/run-test
+++ b/Tests/EndToEndTests/CNTKv2Library/UnitTests/run-test
@@ -44,6 +44,20 @@ DeleteModelsAfterTest=0
 cntkrun 01_OneHidden.cntk "stderr=- command=trainNetwork trainNetwork=[SGD=[maxEpochs=1]]" || exit $?
 cp $OutputDir/Models/01_OneHidden $TestDataDir || exit $?
 
+
+# Train model that is used by LoadLegacyModel tests.
+# Setup predefined variables $DataDir, $ConfigDir, and $OutputDir that are required by cntkrun
+ConfigDir=$TEST_ROOT_DIR/../../Examples/Speech/AN4/Config
+OutputDir=$ConfigDir/../Output
+DataDir=$ConfigDir/../Data
+
+[ -d $OutputDir ] || mkdir $OutputDir || exit $?
+[ -d $OutputDir/Models ] && rm -rf $OutputDir/Models
+DeleteModelsAfterTest=0
+[ -f $ConfigDir/FeedForward.cntk ] || exit 1
+cntkrun FeedForward.cntk "stderr=- command=speechTrain parallelTrain=false speechTrain=[SGD=[maxEpochs=1]]" || exit $?
+cp $OutputDir/Models/cntkSpeechFF.dnn $TestDataDir || exit $?
+
 # Set CUDA_VISIBLE_DEVICES to exclude all gpu if running on cpu device
 if [ "$TEST_DEVICE" == "cpu" ]; then
   export CUDA_VISIBLE_DEVICES=-1
diff --git a/Tests/UnitTests/V2LibraryTests/Common.h b/Tests/UnitTests/V2LibraryTests/Common.h
index c5bda2cbd..f86a24fd8 100644
--- a/Tests/UnitTests/V2LibraryTests/Common.h
+++ b/Tests/UnitTests/V2LibraryTests/Common.h
@@ -640,3 +640,46 @@ inline void CompareFunctions(const FunctionPtr& first, const FunctionPtr& second
         }
     }
 }
+
+inline  MinibatchSourcePtr CreateHTKMinibatchSource(size_t featureDim, size_t numOutputClasses, const Dictionary& readModeConfig, size_t epochSize, bool randomize = true)
+{
+    auto featuresFilePath = L"glob_0000.scp";
+    auto labelsFilePath = L"glob_0000.mlf";
+    auto labelMappingFile = L"state.list";
+
+    Dictionary featuresStreamConfig;
+    featuresStreamConfig[L"dim"] = featureDim;
+    featuresStreamConfig[L"scpFile"] = featuresFilePath;
+
+    CNTK::Dictionary featInputStreamsConfig;
+    featInputStreamsConfig[L"features"] = featuresStreamConfig;
+
+    CNTK::Dictionary featDeserializerConfiguration;
+    featDeserializerConfiguration[L"type"] = L"HTKFeatureDeserializer";
+    featDeserializerConfiguration[L"input"] = featInputStreamsConfig;
+
+    Dictionary labelsStreamConfig;
+    labelsStreamConfig[L"dim"] = numOutputClasses;
+    labelsStreamConfig[L"mlfFile"] = labelsFilePath;
+    labelsStreamConfig[L"labelMappingFile"] = labelMappingFile;
+    labelsStreamConfig[L"scpFile"] = featuresFilePath;
+
+    CNTK::Dictionary labelsInputStreamsConfig;
+    labelsInputStreamsConfig[L"labels"] = labelsStreamConfig;
+
+    CNTK::Dictionary labelsDeserializerConfiguration;
+    labelsDeserializerConfiguration[L"type"] = L"HTKMLFDeserializer";
+    labelsDeserializerConfiguration[L"input"] = labelsInputStreamsConfig;
+
+    Dictionary minibatchSourceConfiguration;
+    if (randomize)
+        minibatchSourceConfiguration[L"randomize"] = true;
+
+    minibatchSourceConfiguration[L"epochSize"] = epochSize;
+    minibatchSourceConfiguration[L"deserializers"] = std::vector<DictionaryValue>({ featDeserializerConfiguration, labelsDeserializerConfiguration });
+    minibatchSourceConfiguration.Add(readModeConfig);
+
+    return CreateCompositeMinibatchSource(minibatchSourceConfiguration);
+}
+
+
diff --git a/Tests/UnitTests/V2LibraryTests/LoadLegacyModelTests.cpp b/Tests/UnitTests/V2LibraryTests/LoadLegacyModelTests.cpp
new file mode 100644
index 000000000..54d545e7c
--- /dev/null
+++ b/Tests/UnitTests/V2LibraryTests/LoadLegacyModelTests.cpp
@@ -0,0 +1,73 @@
+//
+// Copyright (c) Microsoft. All rights reserved.
+// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
+//
+// EvalMultithreads.cpp : Sample application shows how to evaluate a model in multiple threading environment. 
+//
+#include <functional>
+#include <thread>
+#include <iostream>
+#include "CNTKLibrary.h"
+#include "Common.h"
+
+using namespace CNTK;
+
+void TestLoadLegacyModelWithPrecompute(const DeviceDescriptor& device)
+{
+    const size_t baseFeaturesDim = 363;
+    const size_t numOutputClasses = 132;
+
+    auto modelFuncPtr = CNTK::Function::LoadModel(L"cntkSpeechFF.dnn", device);
+
+    auto FindVariableByName = [](const std::vector<Variable>& variables, const std::wstring& name) {
+        for (size_t i = 0; i < variables.size(); ++i)
+            if (variables[i].Name() == name)
+                return variables[i];
+
+        throw std::runtime_error("No output foudn with teh given name");
+    };
+
+    auto arguments = modelFuncPtr->Arguments();
+    auto features = FindVariableByName(arguments, L"features");
+    auto labels = FindVariableByName(arguments, L"labels");
+
+    auto outputs = modelFuncPtr->Outputs();
+    FunctionPtr prediction = FindVariableByName(outputs, L"PosteriorProb");
+    FunctionPtr loss = FindVariableByName(outputs, L"CrossEntropyWithSoftmax");
+    FunctionPtr eval = FindVariableByName(outputs, L"EvalClassificationError");
+
+    Dictionary frameModeConfig;
+    frameModeConfig[L"frameMode"] = true;
+    auto minibatchSource = CreateHTKMinibatchSource(baseFeaturesDim, numOutputClasses, frameModeConfig, MinibatchSource::InfinitelyRepeat, true);
+
+    const size_t minbatchSize = 256;
+    size_t numMinibatches = 10;
+
+    auto featureStreamInfo = minibatchSource->StreamInfo(L"features");
+    auto labelStreamInfo = minibatchSource->StreamInfo(L"labels");
+
+    LearningRatePerSampleSchedule learningRatePerSample = 0.000781;
+    MomentumAsTimeConstantSchedule momentumTimeConstant = 6074;
+    auto learner = MomentumSGDLearner(prediction->Parameters(), learningRatePerSample, momentumTimeConstant, /*unitGainMomentum = */true);
+    auto trainer = CreateTrainer(prediction, loss, eval, { learner });
+
+    for (size_t i = 0; i < numMinibatches; i++)
+    {
+        auto minibatchData = minibatchSource->GetNextMinibatch(minbatchSize, device);
+        if (minibatchData.empty())
+            break;
+
+        trainer->TrainMinibatch({ { features, minibatchData[featureStreamInfo] },{ labels, minibatchData[labelStreamInfo] } }, device);
+        PrintTrainingProgress(trainer, i, 1);
+    }
+}
+
+void LoadLegacyModelTests()
+{
+    fprintf(stderr, "\nLoadLegacyModelTests..\n");
+
+    TestLoadLegacyModelWithPrecompute(DeviceDescriptor::CPUDevice());
+
+    if (IsGPUAvailable())
+        TestLoadLegacyModelWithPrecompute(DeviceDescriptor::GPUDevice(0));
+}
diff --git a/Tests/UnitTests/V2LibraryTests/Main.cpp b/Tests/UnitTests/V2LibraryTests/Main.cpp
index d6ffa5703..c68da9aeb 100644
--- a/Tests/UnitTests/V2LibraryTests/Main.cpp
+++ b/Tests/UnitTests/V2LibraryTests/Main.cpp
@@ -27,6 +27,7 @@ void DeviceSelectionTests();
 void MultiThreadsEvaluation(bool);
 void MinibatchSourceTests();
 void UserDefinedFunctionTests();
+void LoadLegacyModelTests();
 
 int main()
 {
@@ -60,10 +61,11 @@ int main()
 
     FeedForwardTests();
     RecurrentFunctionTests();
-
     UserDefinedFunctionTests();
 
     SerializationTests();
+    LoadLegacyModelTests();
+
     LearnerTests();
 
     TrainerTests();
diff --git a/Tests/UnitTests/V2LibraryTests/TruncatedLSTMAcousticModel.cpp b/Tests/UnitTests/V2LibraryTests/TruncatedLSTMAcousticModel.cpp
index 0465edb11..7eca46da1 100644
--- a/Tests/UnitTests/V2LibraryTests/TruncatedLSTMAcousticModel.cpp
+++ b/Tests/UnitTests/V2LibraryTests/TruncatedLSTMAcousticModel.cpp
@@ -6,47 +6,6 @@ using namespace CNTK;
 
 using namespace std::placeholders;
 
-MinibatchSourcePtr CreateMinibatchSource(size_t featureDim, size_t numOutputClasses, const Dictionary& readModeConfig, size_t epochSize, bool randomize = true)
-{
-    auto featuresFilePath = L"glob_0000.scp";
-    auto labelsFilePath = L"glob_0000.mlf";
-    auto labelMappingFile = L"state.list";
-
-    Dictionary featuresStreamConfig;
-    featuresStreamConfig[L"dim"] = featureDim;
-    featuresStreamConfig[L"scpFile"] = featuresFilePath;
-
-    CNTK::Dictionary featInputStreamsConfig;
-    featInputStreamsConfig[L"features"] = featuresStreamConfig;
-
-    CNTK::Dictionary featDeserializerConfiguration;
-    featDeserializerConfiguration[L"type"] = L"HTKFeatureDeserializer";
-    featDeserializerConfiguration[L"input"] = featInputStreamsConfig;
-
-    Dictionary labelsStreamConfig;
-    labelsStreamConfig[L"dim"] = numOutputClasses;
-    labelsStreamConfig[L"mlfFile"] = labelsFilePath;
-    labelsStreamConfig[L"labelMappingFile"] = labelMappingFile;
-    labelsStreamConfig[L"scpFile"] = featuresFilePath;
-
-    CNTK::Dictionary labelsInputStreamsConfig;
-    labelsInputStreamsConfig[L"labels"] = labelsStreamConfig;
-
-    CNTK::Dictionary labelsDeserializerConfiguration;
-    labelsDeserializerConfiguration[L"type"] = L"HTKMLFDeserializer";
-    labelsDeserializerConfiguration[L"input"] = labelsInputStreamsConfig;
-
-    Dictionary minibatchSourceConfiguration;
-    if (randomize)
-        minibatchSourceConfiguration[L"randomize"] = true;
-
-    minibatchSourceConfiguration[L"epochSize"] = epochSize;
-    minibatchSourceConfiguration[L"deserializers"] = std::vector<DictionaryValue>({ featDeserializerConfiguration, labelsDeserializerConfiguration });
-    minibatchSourceConfiguration.Add(readModeConfig);
-
-    return CreateCompositeMinibatchSource(minibatchSourceConfiguration);
-}
-
 static FunctionPtr LSTMAcousticSequenceClassiferNet(const Variable& input, size_t numOutputClasses, size_t LSTMDim, size_t cellDim, size_t numLSTMs, const DeviceDescriptor& device, const std::wstring& outputName)
 {
     auto pastValueRecurrenceHook = [](const Variable& x) { return PastValue(x); };
@@ -71,7 +30,7 @@ void TrainTruncatedLSTMAcousticModelClassifer(const DeviceDescriptor& device, bo
     const size_t numSamplesForFeatureStatistics = MinibatchSource::FullDataSweep;
     Dictionary frameModeConfig;
     frameModeConfig[L"frameMode"] = true;
-    auto minibatchSource = CreateMinibatchSource(baseFeaturesDim, numOutputClasses, frameModeConfig, numSamplesForFeatureStatistics, false);
+    auto minibatchSource = CreateHTKMinibatchSource(baseFeaturesDim, numOutputClasses, frameModeConfig, numSamplesForFeatureStatistics, false);
     auto featureStreamInfo = minibatchSource->StreamInfo(features);
     std::unordered_map<StreamInformation, std::pair<NDArrayViewPtr, NDArrayViewPtr>> featureMeansAndInvStdDevs = { { featureStreamInfo, { nullptr, nullptr } } };
     ComputeInputPerDimMeansAndInvStdDevs(minibatchSource, featureMeansAndInvStdDevs);
@@ -100,7 +59,7 @@ void TrainTruncatedLSTMAcousticModelClassifer(const DeviceDescriptor& device, bo
     Dictionary truncatedModeConfig;
     truncatedModeConfig[L"truncated"] = true;
     truncatedModeConfig[L"truncationLength"] = truncationLength;
-    minibatchSource = CreateMinibatchSource(baseFeaturesDim, numOutputClasses, truncatedModeConfig, numTrainingSamples);
+    minibatchSource = CreateHTKMinibatchSource(baseFeaturesDim, numOutputClasses, truncatedModeConfig, numTrainingSamples);
 
     const size_t numberParallelSequencesPerMB1 = 16;
     const size_t numberParallelSequencesPerMB2 = 32;
diff --git a/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj b/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj
index cb6dcacb8..f358cc3aa 100644
--- a/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj
+++ b/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj
@@ -120,6 +120,7 @@
     <ClCompile Include="Common.cpp" />
     <ClCompile Include="DeviceSelectionTests.cpp" />
     <ClCompile Include="LearnerTests.cpp" />
+    <ClCompile Include="LoadLegacyModelTests.cpp" />
     <ClCompile Include="MinibatchSourceTest.cpp" />
     <ClCompile Include="Seq2Seq.cpp" />
     <ClCompile Include="SerializationTests.cpp" />
diff --git a/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj.filters b/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj.filters
index 300bbbc0a..471370587 100644
--- a/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj.filters
+++ b/Tests/UnitTests/V2LibraryTests/V2LibraryTests.vcxproj.filters
@@ -78,6 +78,9 @@
     <ClCompile Include="..\..\..\Examples\Evaluation\CNTKLibraryCPPEvalExamples\EvalMultithreads.cpp">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="LoadLegacyModelTests.cpp">
+      <Filter>Source Files</Filter>
+    </ClCompile>
   </ItemGroup>
   <ItemGroup>
     <ClInclude Include="Common.h">

From 1326a44dd8bce0c2e0c33fde8aadc76652dfcf55 Mon Sep 17 00:00:00 2001
From: Eldar Akchurin <eldak@microsoft.com>
Date: Wed, 1 Feb 2017 17:19:39 +0100
Subject: [PATCH 5/7] Adding diagnostics of std::exception

---
 bindings/python/cntk/cntk_py.i | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/bindings/python/cntk/cntk_py.i b/bindings/python/cntk/cntk_py.i
index b04fe9a22..16e33caf2 100644
--- a/bindings/python/cntk/cntk_py.i
+++ b/bindings/python/cntk/cntk_py.i
@@ -581,10 +581,11 @@ public:
 //
 %exception {
     try { $action }
-    catch (Swig::DirectorException &e) { SWIG_exception(SWIG_RuntimeError,e.what()); }
-    catch (std::runtime_error &e) { SWIG_exception(SWIG_RuntimeError,e.what()); }
-    catch (std::invalid_argument &e) { SWIG_exception(SWIG_ValueError,e.what()); }
-    catch (std::logic_error &e) { SWIG_exception(SWIG_RuntimeError,e.what()); }
+    catch (const Swig::DirectorException &e) { SWIG_exception(SWIG_RuntimeError, e.what()); }
+    catch (const std::runtime_error &e) { SWIG_exception(SWIG_RuntimeError, e.what()); }
+    catch (const std::invalid_argument &e) { SWIG_exception(SWIG_ValueError, e.what()); }
+    catch (const std::logic_error &e) { SWIG_exception(SWIG_RuntimeError, e.what()); }
+    catch (const std::exception &e) { SWIG_exception(SWIG_UnknownError, e.what()); }
     catch (...) { SWIG_exception(SWIG_UnknownError,"Runtime exception"); }
 }
 

From 5d2495b46fe1752b1841564113764e6bc0bb69ca Mon Sep 17 00:00:00 2001
From: Eldar Akchurin <eldak@microsoft.com>
Date: Wed, 1 Feb 2017 11:44:51 +0100
Subject: [PATCH 6/7] Fixing inferrence of outputs to allow crossing of dll
 boundary

---
 Source/CNTKv2LibraryDll/API/CNTKLibrary.h     | 12 +++++-
 Source/CNTKv2LibraryDll/BlockFunction.h       |  7 +---
 Source/CNTKv2LibraryDll/CompositeFunction.h   |  5 ++-
 Source/CNTKv2LibraryDll/Function.cpp          | 18 ++++++---
 Source/CNTKv2LibraryDll/PrimitiveFunction.cpp |  9 +++--
 Source/CNTKv2LibraryDll/PrimitiveFunction.h   |  2 +-
 .../UserDefinedFunctionTests.cpp              |  5 +--
 bindings/python/cntk/cntk_py.i                | 38 +++++++++++++++++++
 bindings/python/cntk/ops/functions.py         |  3 ++
 9 files changed, 77 insertions(+), 22 deletions(-)

diff --git a/Source/CNTKv2LibraryDll/API/CNTKLibrary.h b/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
index ae693e6bc..ef132b8a2 100644
--- a/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
+++ b/Source/CNTKv2LibraryDll/API/CNTKLibrary.h
@@ -2628,8 +2628,11 @@ namespace CNTK
         ///
         /// Infers the shape, data type and dynamic axes of the outputs of 'this' function based on the 
         /// Function's inputs, and returns Output Variable objects containing the inferred information
+        /// Result cannot exceed the max number of outputs (128).
+        /// The passed "outputs" vector should also reserve 128 elements in order to not cause memory allocation during
+        /// crossing of dll boundary.
         ///
-        CNTK_API virtual std::vector<Variable> InferOutputs() = 0;
+        CNTK_API virtual void InferOutputs(std::vector<Variable>& outputs) = 0;
 
     public:
 
@@ -2834,6 +2837,11 @@ namespace CNTK
         ///
         CNTK_API void PrintGraph() const;
 
+        ///
+        /// Maimum number of outputs that is currently supported.
+        ///
+        static const int MaxNumOutputs = 64;
+
     protected:
         ///
         /// Protected constructor for derived 'Function' types to specify the actual input and output variables for the (primitive) Function instance.
@@ -2883,7 +2891,7 @@ namespace CNTK
         CNTK_API std::shared_ptr<std::vector<Variable>> InputsImpl(bool pythonOperandOrder = false) const;
         CNTK_API std::shared_ptr<std::vector<Variable>> OutputsImpl() const;
 
-        void ValidateOrUpdateOutputs(std::unordered_map<const Function*, size_t>& visitedFunctions, bool& recurrentNodeOutputModified);
+        void ValidateOrUpdateOutputs(std::unordered_map<const Function*, size_t>& visitedFunctions, bool& recurrentNodeOutputModified, std::vector<Variable>& buffer);
 
         static void ReplacePlaceholderInPlace(Variable& var,
                                               const std::unordered_map<Variable, Variable>& placeholderReplacements,
diff --git a/Source/CNTKv2LibraryDll/BlockFunction.h b/Source/CNTKv2LibraryDll/BlockFunction.h
index 5df80da78..ff3d5ebb3 100644
--- a/Source/CNTKv2LibraryDll/BlockFunction.h
+++ b/Source/CNTKv2LibraryDll/BlockFunction.h
@@ -131,7 +131,7 @@ namespace CNTK
             return blockFunctionInputs;
         }
 
-        virtual std::vector<Variable> InferOutputs() override
+        void InferOutputs(std::vector<Variable>& outputs) override
         {
             // We determine the outputs by replacing the arguments of the composite with new placeholders with updated 
             // shape etc. information matching the corresponding mapped input
@@ -148,17 +148,14 @@ namespace CNTK
 
             m_composite->ReplacePlaceholders(replacementMap);
 
-            std::vector<Variable> blockFunctionOutputs;
             auto compositeOutputs = m_composite->RawOutputs();
             for (auto compositeOutput : compositeOutputs)
             {
                 auto output = OutputVariable(compositeOutput.Shape(), compositeOutput.GetDataType(), compositeOutput.DynamicAxes(), Name());
                 output.m_dataFields->m_blockFunctionVariableMapping = compositeOutput;
 
-                blockFunctionOutputs.push_back(output);
+                outputs.push_back(output);
             }
-
-            return blockFunctionOutputs;
         }
 
     private:
diff --git a/Source/CNTKv2LibraryDll/CompositeFunction.h b/Source/CNTKv2LibraryDll/CompositeFunction.h
index 74a29facf..7dc054637 100644
--- a/Source/CNTKv2LibraryDll/CompositeFunction.h
+++ b/Source/CNTKv2LibraryDll/CompositeFunction.h
@@ -93,9 +93,10 @@ namespace CNTK
             NOT_IMPLEMENTED;
         }
 
-        virtual std::vector<Variable> InferOutputs() override
+        void InferOutputs(std::vector<Variable>& outputs) override
         {
-            return m_rootFunction->InitOutputs();
+            auto& inferred = m_rootFunction->InitOutputs();
+            outputs.assign(inferred.begin(), inferred.end());
         }
 
         virtual void Backward(const BackPropStatePtr& state,
diff --git a/Source/CNTKv2LibraryDll/Function.cpp b/Source/CNTKv2LibraryDll/Function.cpp
index 974fe1de9..144c37025 100644
--- a/Source/CNTKv2LibraryDll/Function.cpp
+++ b/Source/CNTKv2LibraryDll/Function.cpp
@@ -16,7 +16,9 @@ namespace CNTK
     std::vector<Variable>& Function::InitOutputs()
     {
         std::call_once(m_outputsInitFlag, [this]() {
-            auto outputs = InferOutputs();
+            std::vector<Variable> outputs;
+            outputs.reserve(Function::MaxNumOutputs);
+            InferOutputs(outputs);
             std::unordered_set<Variable> uniqueOutputs;
             for (auto outputVar : outputs)
             {
@@ -301,7 +303,7 @@ namespace CNTK
         return updated;
     }
 
-    void Function::ValidateOrUpdateOutputs(std::unordered_map<const Function*, size_t>& visitedFunctions, bool& recurrentNodeOutputModified)
+    void Function::ValidateOrUpdateOutputs(std::unordered_map<const Function*, size_t>& visitedFunctions, bool& recurrentNodeOutputModified, std::vector<Variable>& outputsUsingNewInputs)
     {
         assert(visitedFunctions.find(this) == visitedFunctions.end());
         visitedFunctions[this] = 1;
@@ -313,13 +315,17 @@ namespace CNTK
             {
                 auto owner = input.Owner().get();
                 if (visitedFunctions.find(owner) == visitedFunctions.end())
-                    owner->ValidateOrUpdateOutputs(visitedFunctions, recurrentNodeOutputModified);
+                {
+                    outputsUsingNewInputs.clear();
+                    owner->ValidateOrUpdateOutputs(visitedFunctions, recurrentNodeOutputModified, outputsUsingNewInputs);
+                }
                 else
                     visitedFunctions[owner]++;
             }
         }
 
-        auto outputsUsingNewInputs = this->InferOutputs();
+        outputsUsingNewInputs.clear();
+        this->InferOutputs(outputsUsingNewInputs);
         auto currentOutputs = RawOutputs();
         for (size_t i = 0; i < currentOutputs.size(); ++i)
         {
@@ -480,11 +486,13 @@ namespace CNTK
         const size_t maxNumValidationPassesAllowed = 128;
         bool recurrentNodeOutputModified = false;
         size_t numValidationPasses = 0;
+        std::vector<Variable> outputVarBuffer;
+        outputVarBuffer.reserve(Function::MaxNumOutputs);
         do
         {
             recurrentNodeOutputModified = false;
             functionVisitCounts.clear();
-            RootFunction()->ValidateOrUpdateOutputs(functionVisitCounts, recurrentNodeOutputModified);
+            RootFunction()->ValidateOrUpdateOutputs(functionVisitCounts, recurrentNodeOutputModified, outputVarBuffer);
             numValidationPasses++;
         } while (recurrentNodeOutputModified && (numValidationPasses < maxNumValidationPassesAllowed));
 
diff --git a/Source/CNTKv2LibraryDll/PrimitiveFunction.cpp b/Source/CNTKv2LibraryDll/PrimitiveFunction.cpp
index fb224a993..a8e89e066 100644
--- a/Source/CNTKv2LibraryDll/PrimitiveFunction.cpp
+++ b/Source/CNTKv2LibraryDll/PrimitiveFunction.cpp
@@ -214,10 +214,13 @@ namespace CNTK
         return outputDynamicAxes;
     }
 
-    /*virtual*/ std::vector<Variable> PrimitiveFunction::InferOutputs() /*override*/
+    void PrimitiveFunction::InferOutputs(std::vector<Variable>& outputs)
     {
         if (m_op == PrimitiveOpType::Combine)
-            return m_inputs;
+        {
+            outputs.assign(m_inputs.begin(), m_inputs.end());
+            return;
+        }
 
         DataType outputDataType = GetOutputDataType(m_op, m_inputs, true);
         std::vector<Axis> outputDynamicAxes = GetOutputDynamicAxes(m_op, m_inputs, m_attributes);
@@ -680,7 +683,7 @@ namespace CNTK
             }
         }
 
-        return{ OutputVariable(outputShape, outputDataType, outputDynamicAxes, Name().empty() ? L"" : Name()) };
+        outputs.push_back({ OutputVariable(outputShape, outputDataType, outputDynamicAxes, Name().empty() ? L"" : Name()) });
     }
 
     static const std::wstring s_primitiveFunctionTypeValue = L"PrimitiveFunction";
diff --git a/Source/CNTKv2LibraryDll/PrimitiveFunction.h b/Source/CNTKv2LibraryDll/PrimitiveFunction.h
index 18b16c96d..ea17c6a37 100644
--- a/Source/CNTKv2LibraryDll/PrimitiveFunction.h
+++ b/Source/CNTKv2LibraryDll/PrimitiveFunction.h
@@ -699,7 +699,7 @@ namespace CNTK
         static DataType GetOutputDataType(PrimitiveOpType op, std::vector<Variable>& inputs, bool inferDimensions);
         static std::vector<Axis> GetOutputDynamicAxes(PrimitiveOpType op, std::vector<Variable>& inputs, Dictionary& functionConfig);
 
-        virtual std::vector<Variable> InferOutputs() override;
+        void InferOutputs(std::vector<Variable>& outputs) override;
 
     private:
         PrimitiveOpType m_op;
diff --git a/Tests/UnitTests/V2LibraryTests/UserDefinedFunctionTests.cpp b/Tests/UnitTests/V2LibraryTests/UserDefinedFunctionTests.cpp
index 18298d81d..79e4a918f 100644
--- a/Tests/UnitTests/V2LibraryTests/UserDefinedFunctionTests.cpp
+++ b/Tests/UnitTests/V2LibraryTests/UserDefinedFunctionTests.cpp
@@ -110,18 +110,15 @@ public:
     size_t CurrentVersion() const override { NOT_IMPLEMENTED; }
 
 private:
-    std::vector<Variable> InferOutputs() override
+    void InferOutputs(std::vector<Variable>& outputs) override
     {
         auto leftOperand = Inputs()[0];
         auto rightOperand = Inputs()[1];
         auto tempFunc = m_isTimes ? Times(leftOperand, rightOperand) : Plus(leftOperand, rightOperand);
         auto tempFuncOutputs = tempFunc->Outputs();
 
-        std::vector<Variable> outputs;
         for (auto tempFuncOutput : tempFuncOutputs)
             outputs.push_back(OutputVariable(tempFuncOutput.Shape(), tempFuncOutput.GetDataType(), tempFuncOutput.DynamicAxes()));
-
-        return outputs;
     }
 
     UserDefinedTimesOrPlusFunction(const Variable& leftOperand, const Variable& rightOperand, bool isTimes, const std::wstring& name)
diff --git a/bindings/python/cntk/cntk_py.i b/bindings/python/cntk/cntk_py.i
index 16e33caf2..5cc87a0d3 100644
--- a/bindings/python/cntk/cntk_py.i
+++ b/bindings/python/cntk/cntk_py.i
@@ -17,6 +17,7 @@
 
 %rename(_forward) CNTK::Function::Forward;
 %rename(_backward) CNTK::Function::Backward;
+%rename(_infer_outputs) CNTK::Function::InferOutputs;
 %rename(sgd_learner) CNTK::SGDLearner;
 %rename(momentum_sgd_learner) CNTK::MomentumSGDLearner;
 %rename(gpu_device) CNTK::DeviceDescriptor::GPUDevice;
@@ -740,6 +741,43 @@ public:
 %enddef
 
 
+// Implementing typemapping for InferOutputs virtual function that takes vector of variables by reference
+// And can be implemented in Python.
+
+%typemap(directorin) std::vector<CNTK::Variable>& outputs
+{
+    $input = PyList_New(0);
+}
+
+%typemap(directorargout) std::vector<CNTK::Variable>& outputs
+{
+    if (!PyList_Check($input))
+        RuntimeError("List expected");
+
+    auto iterator = std::shared_ptr<PyObject>(PyObject_GetIter($input), [](PyObject* p) { Py_DECREF(p); });
+    if (!iterator)
+        RuntimeError("Cannot get iterator to the list");
+
+     PyObject *item = nullptr;
+     while ((item = PyIter_Next(iterator.get())))
+     {
+         void *raw_var = nullptr;
+         int res = SWIG_ConvertPtr(item, &raw_var, swig::type_info<CNTK::Variable>(),  0);
+         if (!SWIG_IsOK(res))
+             RuntimeError("Cannot convert list element to CNTK::Variable");
+
+         if (!raw_var)
+             RuntimeError("Invalid null reference when converting a list element to CNTK::Variable");
+
+         auto var = reinterpret_cast<CNTK::Variable*>(raw_var);
+         $1.push_back(*var);
+         Py_DECREF(item);
+     }
+
+     if (PyErr_Occurred())
+         RuntimeError("Cannot convert list element to CNTK::Variable");
+}
+
 // For the output dict (the non-const unordered_map) we need to get the
 // modified values and put them back into the dictionary. This is used, when
 // e.g. the user puts a variable into the dictionary, hoping that it will
diff --git a/bindings/python/cntk/ops/functions.py b/bindings/python/cntk/ops/functions.py
index 189f8f896..397f42f55 100644
--- a/bindings/python/cntk/ops/functions.py
+++ b/bindings/python/cntk/ops/functions.py
@@ -749,6 +749,9 @@ class UserFunction(Function):
 
             variables[k] = sanitize_batch(k, v, None, state.device())
 
+    def _infer_outputs(self, outputs):
+        outputs.extend(self.infer_outputs())
+
     def infer_outputs(self):
         raise NotImplementedError('infer_outputs has to be overridden')
 

From e4597c94612d51ed8d0f3d0d01540173f0834533 Mon Sep 17 00:00:00 2001
From: Zhou Wang <zhouwang@microsoft.com>
Date: Thu, 2 Feb 2017 14:59:35 +0100
Subject: [PATCH 7/7] change reference type from project reference to file
 refernce to make msbuild happy; add prect dependency

---
 CNTK.sln                                                   | 3 +++
 .../CNTKLibraryCSEvalExamplesTest.csproj                   | 7 +------
 2 files changed, 4 insertions(+), 6 deletions(-)

diff --git a/CNTK.sln b/CNTK.sln
index ecd82e9de..cb1846746 100644
--- a/CNTK.sln
+++ b/CNTK.sln
@@ -1452,6 +1452,9 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "PerformanceProfilerDll", "S
 	EndProjectSection
 EndProject
 Project("{FAE04EC0-301F-11D3-BF4B-00C04F79EFBC}") = "CNTKLibraryCSEvalExamplesTest", "Tests\EndToEndTests\EvalClientTests\CNTKLibraryCSEvalExamplesTest\CNTKLibraryCSEvalExamplesTest.csproj", "{3500A847-E024-4E7D-92DD-CC587C17460B}"
+	ProjectSection(ProjectDependencies) = postProject
+		{50EF9EE6-5018-453E-A063-F77044EF1A97} = {50EF9EE6-5018-453E-A063-F77044EF1A97}
+	EndProjectSection
 EndProject
 Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "GoogLeNet", "GoogLeNet", "{789B4AB8-40F1-4A37-823A-BC20D80C8BF1}"
 	ProjectSection(SolutionItems) = preProject
diff --git a/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj b/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj
index 35ac9485a..0bc03b943 100644
--- a/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj
+++ b/Tests/EndToEndTests/EvalClientTests/CNTKLibraryCSEvalExamplesTest/CNTKLibraryCSEvalExamplesTest.csproj
@@ -60,6 +60,7 @@
     <AllowUnsafeBlocks>false</AllowUnsafeBlocks>
   </PropertyGroup>
   <ItemGroup>
+    <Reference Include="CNTKLibraryManaged-2.0, Version=1.0.0.0, Culture=neutral, processorArchitecture=AMD64" />
     <Reference Include="System" />
     <Reference Include="System.Drawing" />
   </ItemGroup>
@@ -69,12 +70,6 @@
   <ItemGroup>
     <None Include="App.config" />
   </ItemGroup>
-  <ItemGroup>
-    <ProjectReference Include="..\..\..\..\bindings\csharp\CNTKLibraryManagedDll\CNTKLibraryManagedDll.csproj">
-      <Project>{50ef9ee6-5018-453e-a063-f77044ef1a97}</Project>
-      <Name>CNTKLibraryManagedDll</Name>
-    </ProjectReference>
-  </ItemGroup>
   <ItemGroup>
     <Compile Include="..\..\..\..\Examples\Evaluation\CNTKLibraryCSEvalCPUOnlyExamples\CntkBitmapExtensions.cs">
       <Link>CntkBitmapExtensions.cs</Link>