microsoft
/
CNTK
зеркало из
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

CNTKv2Library: Linux port

This commit is contained in:
Amit 2016-06-11 12:21:15 -07:00
Родитель c508e9f4ca
Коммит 26c2006541
16 изменённых файлов: 301 добавлений и 214 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

122
Makefile
Просмотреть файл

@ -60,7 +60,7 @@ endif
CXX = mpic++
SOURCEDIR:= Source
INCLUDEPATH:= $(addprefix $(SOURCEDIR)/, Common/Include Math CNTK ActionsLib ComputationNetworkLib SGDLib SequenceTrainingLib CNTK/BrainScript Readers/ReaderLib)
INCLUDEPATH:= $(addprefix $(SOURCEDIR)/, Common/Include CNTKv2LibraryDll CNTKv2LibraryDll/API Math CNTK ActionsLib ComputationNetworkLib SGDLib SequenceTrainingLib CNTK/BrainScript Readers/ReaderLib)
# COMMON_FLAGS include settings that are passed both to NVCC and C++ compilers.
COMMON_FLAGS:= -D_POSIX_SOURCE -D_XOPEN_SOURCE=600 -D__USE_XOPEN2K -std=c++11
CPPFLAGS:=
@ -247,7 +247,7 @@ READER_SRC =\
$(SOURCEDIR)/Readers/ReaderLib/TruncatedBpttPacker.cpp \
$(SOURCEDIR)/Readers/ReaderLib/PackerBase.cpp \
$(SOURCEDIR)/Readers/ReaderLib/FramePacker.cpp \
$(SOURCEDIR)/Readers/ReaderLib/ChunkCache.cpp \
$(SOURCEDIR)/Readers/ReaderLib/ChunkCache.cpp \
COMMON_SRC =\
$(SOURCEDIR)/Common/Config.cpp \
@ -309,7 +309,96 @@ $(CNTKMATH_LIB): $(MATH_OBJ)
@mkdir -p $(dir $@)
$(CXX) $(LDFLAGS) -shared $(patsubst %,-L%, $(LIBPATH) $(NVMLPATH)) $(patsubst %,$(RPATH)%, $(ORIGINDIR) $(LIBPATH)) -o $@ $^ $(LIBS) -fopenmp
# CNTKLibrary
########################################
CNTK_COMMON_SRC =\
$(SOURCEDIR)/Common/BestGpu.cpp \
$(SOURCEDIR)/Common/MPIWrapper.cpp \
COMPUTATION_NETWORK_LIB_SRC =\
$(SOURCEDIR)/ComputationNetworkLib/ComputationNode.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNodeScripting.cpp \
$(SOURCEDIR)/ComputationNetworkLib/InputAndParamNodes.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ReshapingNodes.cpp \
$(SOURCEDIR)/ComputationNetworkLib/SpecialPurposeNodes.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetwork.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkEvaluation.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkAnalysis.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkEditing.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkBuilder.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkScripting.cpp \
SEQUENCE_TRAINING_LIB_SRC =\
$(SOURCEDIR)/SequenceTrainingLib/latticeforwardbackward.cpp \
$(SOURCEDIR)/SequenceTrainingLib/parallelforwardbackward.cpp \
ifdef CUDA_PATH
SEQUENCE_TRAINING_LIB_SRC +=\
$(SOURCEDIR)/Math/cudalatticeops.cu \
$(SOURCEDIR)/Math/cudalattice.cpp \
$(SOURCEDIR)/Math/cudalib.cpp \
else
SEQUENCE_TRAINING_LIB_SRC +=\
$(SOURCEDIR)/SequenceTrainingLib/latticeNoGPU.cpp \
endif
CNTKLIBRARY_SRC =\
$(SOURCEDIR)/CNTKv2LibraryDll/Common.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/Function.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/NDArrayView.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/NDMask.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/Utils.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/Value.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/Variable.cpp \
CNTKLIBRARY_SRC+=$(CNTK_COMMON_SRC)
CNTKLIBRARY_SRC+=$(COMPUTATION_NETWORK_LIB_SRC)
CNTKLIBRARY_SRC+=$(SEQUENCE_TRAINING_LIB_SRC)
CNTKLIBRARY_VERSION=2.0
CNTKLIBRARY:=cntklibrary-$(CNTKLIBRARY_VERSION)
CNTKLIBRARY_OBJ := $(patsubst %.cu, $(OBJDIR)/%.o, $(patsubst %.cpp, $(OBJDIR)/%.o, $(CNTKLIBRARY_SRC)))
CNTKLIBRARY_LIB:=$(LIBDIR)/lib$(CNTKLIBRARY).so
ALL+=$(CNTKLIBRARY_LIB)
SRC+=$(CNTKLIBRARY_SRC)
RPATH=-Wl,-rpath,
$(CNTKLIBRARY_LIB): $(CNTKLIBRARY_OBJ) | $(CNTKMATH_LIB)
@echo $(SEPARATOR)
@mkdir -p $(dir $@)
@echo building output for $(ARCH) with build type $(BUILDTYPE)
$(CXX) $(LDFLAGS) -shared $(patsubst %,-L%, $(LIBDIR) $(LIBPATH) $(NVMLPATH)) $(patsubst %,$(RPATH)%, $(ORIGINDIR) $(LIBPATH)) -o $@ $^ $(LIBS) -l$(CNTKMATH)
# CNTKLibrary tests
########################################
CNTKLIBRARY_TESTS_SRC =\
$(SOURCEDIR)/CNTKv2LibraryDll/LibraryTests/FeedForwardTests.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/LibraryTests/Main.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/LibraryTests/NDArrayViewTests.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/LibraryTests/RecurrentFunctionTests.cpp \
$(SOURCEDIR)/CNTKv2LibraryDll/LibraryTests/TensorTests.cpp \
CNTKLIBRARY_TESTS:=$(BINDIR)/cntklibrarytests
CNTKLIBRARY_TESTS_OBJ := $(patsubst %.cu, $(OBJDIR)/%.o, $(patsubst %.cpp, $(OBJDIR)/%.o, $(CNTKLIBRARY_TESTS_SRC)))
ALL+=$(CNTKLIBRARY_TESTS)
SRC+=$(CNTKLIBRARY_TESTS_SRC)
RPATH=-Wl,-rpath,
$(CNTKLIBRARY_TESTS): $(CNTKLIBRARY_TESTS_OBJ) | $(CNTKLIBRARY_LIB)
@echo $(SEPARATOR)
@mkdir -p $(dir $@)
@echo building output for $(ARCH) with build type $(BUILDTYPE)
$(CXX) $(LDFLAGS) $(patsubst %,-L%, $(LIBDIR) $(LIBPATH) $(NVMLPATH)) $(patsubst %,$(RPATH)%, $(ORIGINLIBDIR) $(LIBPATH)) -o $@ $^ $(LIBS) -l$(CNTKLIBRARY) -l$(CNTKMATH)
# BinaryReader plugin
########################################
@ -594,17 +683,6 @@ CNTK_SRC =\
$(SOURCEDIR)/CNTK/CNTK.cpp \
$(SOURCEDIR)/CNTK/ModelEditLanguage.cpp \
$(SOURCEDIR)/CNTK/tests.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNode.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNodeScripting.cpp \
$(SOURCEDIR)/ComputationNetworkLib/InputAndParamNodes.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ReshapingNodes.cpp \
$(SOURCEDIR)/ComputationNetworkLib/SpecialPurposeNodes.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetwork.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkEvaluation.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkAnalysis.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkEditing.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkBuilder.cpp \
$(SOURCEDIR)/ComputationNetworkLib/ComputationNetworkScripting.cpp \
$(SOURCEDIR)/SGDLib/Profiler.cpp \
$(SOURCEDIR)/SGDLib/SGD.cpp \
$(SOURCEDIR)/ActionsLib/TrainActions.cpp \
@ -615,26 +693,14 @@ CNTK_SRC =\
$(SOURCEDIR)/ActionsLib/NetworkDescriptionLanguage.cpp \
$(SOURCEDIR)/ActionsLib/SimpleNetworkBuilder.cpp \
$(SOURCEDIR)/ActionsLib/NDLNetworkBuilder.cpp \
$(SOURCEDIR)/SequenceTrainingLib/latticeforwardbackward.cpp \
$(SOURCEDIR)/SequenceTrainingLib/parallelforwardbackward.cpp \
$(SOURCEDIR)/CNTK/BrainScript/BrainScriptEvaluator.cpp \
$(SOURCEDIR)/CNTK/BrainScript/BrainScriptParser.cpp \
$(SOURCEDIR)/CNTK/BrainScript/BrainScriptTest.cpp \
$(SOURCEDIR)/Common/BestGpu.cpp \
$(SOURCEDIR)/Common/MPIWrapper.cpp \
ifdef CUDA_PATH
CNTK_SRC +=\
$(SOURCEDIR)/Math/cudalatticeops.cu \
$(SOURCEDIR)/Math/cudalattice.cpp \
$(SOURCEDIR)/Math/cudalib.cpp \
else
CNTK_SRC +=\
$(SOURCEDIR)/SequenceTrainingLib/latticeNoGPU.cpp \
endif
CNTK_SRC+=$(CNTK_COMMON_SRC)
CNTK_SRC+=$(COMPUTATION_NETWORK_LIB_SRC)
CNTK_SRC+=$(SEQUENCE_TRAINING_LIB_SRC)
CNTK_OBJ := $(patsubst %.cu, $(OBJDIR)/%.o, $(patsubst %.cpp, $(OBJDIR)/%.o, $(CNTK_SRC)))

44
Source/CNTKv2LibraryDll/API/CNTKLibrary.h
Просмотреть файл

@ -50,7 +50,7 @@ namespace CNTK
/// Get the 'DataType' corresponding to the ElementType template type argument.
///
template <typename ElementType>
inline DataType GetDataType()
inline DataType AsDataType()
{
if (std::is_same<ElementType, float>())
return DataType::Float;
@ -213,9 +213,9 @@ namespace CNTK
///
/// Creates and returns a new NDShape instance with the same dimensions as 'this' shape's specified axis range.
///
NDShape SubShape(size_t startAxisId = 0, size_t endAxisIdExclusive = UINT_MAX) const
NDShape SubShape(size_t startAxisId = 0, size_t endAxisIdExclusive = SIZE_MAX) const
{
endAxisIdExclusive = (endAxisIdExclusive == UINT_MAX) ? NumAxes() : endAxisIdExclusive;
endAxisIdExclusive = (endAxisIdExclusive == SIZE_MAX) ? NumAxes() : endAxisIdExclusive;
if ((endAxisIdExclusive < startAxisId) || (endAxisIdExclusive > NumAxes()))
InvalidArgument("NDShape::SubShape : The specified endAxisId cannot exceed the number of axes of 'this' NDShape and must be >= than the specified startAxisId");
@ -332,7 +332,7 @@ namespace CNTK
///
template <typename ElementType>
NDArrayView(const NDShape& viewShape, ElementType* dataBuffer, size_t numBufferElements, const DeviceDescriptor& device, bool readOnly = false)
: NDArrayView(GetDataType<ElementType>(), viewShape, dataBuffer, numBufferElements * sizeof(ElementType), device, readOnly)
: NDArrayView(AsDataType<ElementType>(), viewShape, dataBuffer, numBufferElements * sizeof(ElementType), device, readOnly)
{}
///
@ -351,7 +351,7 @@ namespace CNTK
///
template <typename ElementType>
explicit NDArrayView(const ElementType& value, const NDShape& viewShape = { 1 }, const DeviceDescriptor& device = DeviceDescriptor::DefaultDevice(), bool readOnly = false)
: NDArrayView(GetDataType<ElementType>(), viewShape, device)
: NDArrayView(AsDataType<ElementType>(), viewShape, device)
{
SetValue(value);
m_isReadOnly = readOnly;
@ -386,7 +386,7 @@ namespace CNTK
///
/// Returns the data type of 'this' view's contents.
///
DataType DataType() const
DataType GetDataType() const
{
return m_dataType;
}
@ -394,7 +394,7 @@ namespace CNTK
///
/// Returns the storage format of 'this' view.
///
StorageFormat StorageFormat() const
StorageFormat GetStorageFormat() const
{
return m_storageFormat;
}
@ -412,7 +412,7 @@ namespace CNTK
///
bool IsSparse() const
{
return (StorageFormat() != StorageFormat::Dense);
return (GetStorageFormat() != StorageFormat::Dense);
}
///
@ -598,14 +598,14 @@ namespace CNTK
/// The created Value object contains a copy of the specified 'sequences' data.
///
template <typename ElementType>
static ValuePtr Create(const NDShape& sampleShape, const std::vector<const std::vector<ElementType>>& sequences, const DeviceDescriptor& device, bool readOnly = false);
static ValuePtr Create(const NDShape& sampleShape, const std::vector<std::vector<ElementType>>& sequences, const DeviceDescriptor& device, bool readOnly = false);
///
/// Create a new Value object containing a collection of variable length sequences of one hot vectors
/// The created Value object contains a copy of the specified 'sequences' data.
///
template <typename ElementType>
static ValuePtr Create(size_t vocabularySize, const std::vector<const std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly = false);
static ValuePtr Create(size_t vocabularySize, const std::vector<std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly = false);
///
/// Destruct 'this' Value object.
@ -960,18 +960,14 @@ namespace CNTK
///
/// Returns the DataType of the data that 'this' Variable symbolically represents
///
DataType DataType() const
DataType GetDataType() const
{
return m_dataFields->m_dataType;
}
#ifdef _WIN32
// BUGBUG: Add a default constructor to workaround a bug in VS2013 std::vector implementation that
// incorrectly requires its elements to be default constructible
Variable()
{
}
#endif
///
/// Returns a boolean value indicating if gradient computation is enabled for this variable.
@ -1057,7 +1053,7 @@ namespace CNTK
/// Construct a parameter whose initial contents are a copy of the specified 'value'
///
explicit Parameter(const NDArrayViewPtr& value, const std::wstring& name = L"")
: Variable(value->Shape(), VariableKind::Parameter, value->DataType(), value->DeepClone(), true, {}, name)
: Variable(value->Shape(), VariableKind::Parameter, value->GetDataType(), value->DeepClone(), true, {}, name)
{
}
@ -1068,7 +1064,7 @@ namespace CNTK
///
template<typename ElemType>
Parameter(const NDShape& shape, ElemType initValue, const DeviceDescriptor& device = DeviceDescriptor::DefaultDevice(), const std::wstring& name = L"")
: Variable(shape, VariableKind::Parameter, GetDataType<ElemType>(), new NDArrayView(initValue, shape, device), true, {}, name)
: Variable(shape, VariableKind::Parameter, AsDataType<ElemType>(), new NDArrayView(initValue, shape, device), true, {}, name)
{
}
@ -1106,7 +1102,7 @@ namespace CNTK
/// Contruct a Constant whose initial contents are a copy of the specified value
///
Constant(const NDArrayViewPtr& value, const std::wstring& name = L"")
: Variable(value->Shape(), VariableKind::Constant, value->DataType(), value->DeepClone(true), false, {}, name)
: Variable(value->Shape(), VariableKind::Constant, value->GetDataType(), value->DeepClone(true), false, {}, name)
{
}
@ -1117,7 +1113,7 @@ namespace CNTK
///
template<typename ElemType>
Constant(const NDShape& shape, ElemType initValue, const DeviceDescriptor& device = DeviceDescriptor::DefaultDevice(), const std::wstring& name = L"")
: Variable(shape, VariableKind::Constant, GetDataType<ElemType>(), new NDArrayView(initValue, shape, device), false, {}, name)
: Variable(shape, VariableKind::Constant, AsDataType<ElemType>(), new NDArrayView(initValue, shape, device), false, {}, name)
{
}
@ -1178,6 +1174,14 @@ namespace CNTK
}
namespace std {
template <> struct hash<CNTK::Axis>
{
size_t operator()(const CNTK::Axis& x) const
{
return std::hash<std::wstring>()(x.Name());
}
};
template <> struct hash<CNTK::Variable>
{
size_t operator()(const CNTK::Variable& x) const
@ -1349,7 +1353,7 @@ namespace CNTK
///
FunctionPtr RootFunction() const
{
return (m_rootFunction == nullptr) ? const_cast<Function*>(this) : m_rootFunction;
return (m_rootFunction == nullptr) ? const_cast<Function*>(this) : m_rootFunction.GetPtr();
}
///

159
Source/CNTKv2LibraryDll/API/CNTKLibraryInternals.h
Просмотреть файл

@ -27,6 +27,8 @@
#include <type_traits>
#include <unordered_set>
#include <unordered_map>
#include <stdlib.h>
#include <string.h>
#pragma warning(disable: 4702 4127)
@ -47,6 +49,83 @@ namespace Microsoft { namespace MSR { namespace CNTK {
class ComputationNode;
}}}
// TODO: The following should be reconciled with the equivalent code in the CNTK implementation
#ifndef _MSC_VER
#define _countof(_Array) (sizeof(_Array) / sizeof(_Array[0]))
#endif
namespace CNTK
{
#define UNUSED(x) (void)(x) // for variables that are, e.g., only used in _DEBUG builds
#ifdef _MSC_VER
#define __declspec_noreturn __declspec(noreturn)
#else
#define __declspec_noreturn __attribute__((noreturn))
#endif
#pragma warning(push)
#pragma warning(disable : 4996)
#ifndef _MSC_VER // TODO: what is the correct trigger for gcc?
template <class E>
__declspec_noreturn void ThrowFormatted(const char* format, ...) __attribute__((format(printf, 1, 2)));
#endif
template <class E>
__declspec_noreturn inline void ThrowFormatted(const char* format, ...)
{
va_list args;
va_start(args, format);
char buffer[1024] = { 0 }; // Note: pre-VS2015 vsnprintf() is not standards-compliant and may not add a terminator
int written = vsnprintf(buffer, _countof(buffer) - 1, format, args); // -1 because pre-VS2015 vsnprintf() does not always write a 0-terminator
// TODO: In case of EILSEQ error, choose between just outputting the raw format itself vs. continuing the half-completed buffer
//if (written < 0) // an invalid wide-string conversion may lead to EILSEQ
// strncpy(buffer, format, _countof(buffer)
UNUSED(written); // pre-VS2015 vsnprintf() returns -1 in case of overflow, instead of the #characters written
if (strlen(buffer)/*written*/ >= (int)_countof(buffer) - 2)
sprintf(buffer + _countof(buffer) - 4, "...");
// TODO: Should use ExceptionWithCallStack; temporarily using std::exception to avoid duplicating headers
//throw ExceptionWithCallStack<E>(buffer, ExceptionWithCallStack<E>::GetCallStack(/*skipLevels=*/2, /*makeFunctionNamesStandOut=*/true));
throw E(buffer);
}
#pragma warning(pop)
// RuntimeError - throw a std::runtime_error with a formatted error string
#ifndef _MSC_VER // gcc __attribute__((format(printf())) does not percolate through variadic templates; so must go the macro route
#define RuntimeError ThrowFormatted<std::runtime_error>
#define LogicError ThrowFormatted<std::logic_error>
#define InvalidArgument ThrowFormatted<std::invalid_argument>
#else
template <class... _Types>
__declspec_noreturn inline void RuntimeError(const char* format, _Types&&... _Args)
{
ThrowFormatted<std::runtime_error>(format, std::forward<_Types>(_Args)...);
}
template <class... _Types>
__declspec_noreturn inline void LogicError(const char* format, _Types&&... _Args)
{
ThrowFormatted<std::logic_error>(format, std::forward<_Types>(_Args)...);
}
template <class... _Types>
__declspec_noreturn inline void InvalidArgument(const char* format, _Types&&... _Args)
{
ThrowFormatted<std::invalid_argument>(format, std::forward<_Types>(_Args)...);
}
#endif
#ifndef NOT_IMPLEMENTED
#define NOT_IMPLEMENTED \
{ \
fprintf(stderr, "Inside File: %s Line: %d Function: %s -> Feature Not Implemented.\n", __FILE__, __LINE__, __FUNCTION__); \
LogicError("Inside File: %s Line: %d Function: %s -> Feature Not Implemented.\n", __FILE__, __LINE__, __FUNCTION__); \
}
#endif
}
namespace CNTK
{
// Forward declarations
@ -234,7 +313,7 @@ namespace CNTK
template <typename ValueType>
friend CNTK_API bool operator==(const _SimpleVector<ValueType>& first, const _SimpleVector<ValueType>& second);
friend class Function;
friend class CNTK::Function;
public:
_SimpleVector();
@ -309,7 +388,7 @@ namespace CNTK
template <typename KeyType>
class CNTK_API _SimpleSet final
{
friend class CompositeFunction;
friend class CNTK::CompositeFunction;
template <typename T>
friend CNTK_API bool operator==(const _SimpleSet<T>& first, const _SimpleSet<T>& second);
@ -363,8 +442,8 @@ namespace CNTK
template <typename KeyType, typename ValueType>
class CNTK_API _SimpleMap final
{
friend class CompositeFunction;
friend class Function;
friend class CNTK::CompositeFunction;
friend class CNTK::Function;
public:
_SimpleMap();
@ -435,75 +514,3 @@ namespace std {
}
};
}
// TODO: The following should be reconciled with the equivalent code in the CNTK implementation
namespace CNTK
{
#define UNUSED(x) (void)(x) // for variables that are, e.g., only used in _DEBUG builds
#ifdef _MSC_VER
#define __declspec_noreturn __declspec(noreturn)
#else
#define __declspec_noreturn __attribute__((noreturn))
#endif
#pragma warning(push)
#pragma warning(disable : 4996)
#ifndef _MSC_VER // TODO: what is the correct trigger for gcc?
template <class E>
__declspec_noreturn void ThrowFormatted(const char* format, ...) __attribute__((format(printf, 1, 2)));
#endif
template <class E>
__declspec_noreturn inline void ThrowFormatted(const char* format, ...)
{
va_list args;
va_start(args, format);
char buffer[1024] = { 0 }; // Note: pre-VS2015 vsnprintf() is not standards-compliant and may not add a terminator
int written = vsnprintf(buffer, _countof(buffer) - 1, format, args); // -1 because pre-VS2015 vsnprintf() does not always write a 0-terminator
// TODO: In case of EILSEQ error, choose between just outputting the raw format itself vs. continuing the half-completed buffer
//if (written < 0) // an invalid wide-string conversion may lead to EILSEQ
// strncpy(buffer, format, _countof(buffer)
UNUSED(written); // pre-VS2015 vsnprintf() returns -1 in case of overflow, instead of the #characters written
if (strlen(buffer)/*written*/ >= (int)_countof(buffer) - 2)
sprintf(buffer + _countof(buffer) - 4, "...");
// TODO: Should use ExceptionWithCallStack; temporarily using std::exception to avoid duplicating headers
//throw ExceptionWithCallStack<E>(buffer, ExceptionWithCallStack<E>::GetCallStack(/*skipLevels=*/2, /*makeFunctionNamesStandOut=*/true));
throw std::exception(buffer);
}
#pragma warning(pop)
// RuntimeError - throw a std::runtime_error with a formatted error string
#ifndef _MSC_VER // gcc __attribute__((format(printf())) does not percolate through variadic templates; so must go the macro route
#define RuntimeError ThrowFormatted<std::runtime_error>
#define LogicError ThrowFormatted<std::logic_error>
#define InvalidArgument ThrowFormatted<std::invalid_argument>
#else
template <class... _Types>
__declspec_noreturn inline void RuntimeError(const char* format, _Types&&... _Args)
{
ThrowFormatted<std::runtime_error>(format, std::forward<_Types>(_Args)...);
}
template <class... _Types>
__declspec_noreturn inline void LogicError(const char* format, _Types&&... _Args)
{
ThrowFormatted<std::logic_error>(format, std::forward<_Types>(_Args)...);
}
template <class... _Types>
__declspec_noreturn inline void InvalidArgument(const char* format, _Types&&... _Args)
{
ThrowFormatted<std::invalid_argument>(format, std::forward<_Types>(_Args)...);
}
#endif
#ifndef NOT_IMPLEMENTED
#define NOT_IMPLEMENTED \
{ \
fprintf(stderr, "Inside File: %s Line: %d Function: %s -> Feature Not Implemented.\n", __FILE__, __LINE__, __FUNCTION__); \
LogicError("Inside File: %s Line: %d Function: %s -> Feature Not Implemented.\n", __FILE__, __LINE__, __FUNCTION__); \
}
#endif
}

4
Source/CNTKv2LibraryDll/CNTKv2LibraryDll.vcxproj
Просмотреть файл

@ -75,7 +75,7 @@
<Link>
<SubSystem>Console</SubSystem>
<GenerateDebugInformation>true</GenerateDebugInformation>
<AdditionalDependencies>ComputationNetworkLib.lib; Math.lib; Common.lib; ActionsLib.lib; kernel32.lib; user32.lib; shell32.lib; SequenceTrainingLib.lib; %(AdditionalDependencies)</AdditionalDependencies>
<AdditionalDependencies>ComputationNetworkLib.lib; Math.lib; Common.lib; SequenceTrainingLib.lib; kernel32.lib; user32.lib; shell32.lib; %(AdditionalDependencies)</AdditionalDependencies>
<DelayLoadDLLs>Math.dll; nvml.dll; $(CudaRuntimeDll)</DelayLoadDLLs>
</Link>
</ItemDefinitionGroup>
@ -99,7 +99,7 @@
<GenerateDebugInformation>true</GenerateDebugInformation>
<EnableCOMDATFolding>true</EnableCOMDATFolding>
<OptimizeReferences>true</OptimizeReferences>
<AdditionalDependencies>ComputationNetworkLib.lib; Math.lib; Common.lib; ActionsLib.lib; kernel32.lib; user32.lib; shell32.lib; SequenceTrainingLib.lib; %(AdditionalDependencies)</AdditionalDependencies>
<AdditionalDependencies>ComputationNetworkLib.lib; Math.lib; Common.lib; kernel32.lib; user32.lib; shell32.lib; SequenceTrainingLib.lib; %(AdditionalDependencies)</AdditionalDependencies>
<Profile>true</Profile>
<DelayLoadDLLs>Math.dll; nvml.dll; $(CudaRuntimeDll)</DelayLoadDLLs>
</Link>

52
Source/CNTKv2LibraryDll/Function.cpp
Просмотреть файл

@ -84,7 +84,7 @@ namespace CNTK
else
{
assert(variable.Kind() == VariableKind::Output);
computationNodePtr = GetOutputVariableNode(variable, network, builder, variableToNodeMap, isVariableRootMap)->As<ComputationNode<ElementType>>()->shared_from_this();
computationNodePtr = GetOutputVariableNode(variable, network, builder, variableToNodeMap, isVariableRootMap)->template As<ComputationNode<ElementType>>()->shared_from_this();
}
variableToNodeMap[variable] = computationNodePtr;
@ -106,13 +106,13 @@ namespace CNTK
// Create the nodes corresponding to the inputs
auto functionInputs = primitiveFunction->Inputs();
auto input0BaseNodePtr = GetNode(functionInputs[0], network, builder, variableToNodeMap, isVariableRootMap);
std::shared_ptr<ComputationNode<ElementType>> input0Node = (input0BaseNodePtr != nullptr) ? input0BaseNodePtr->As<ComputationNode<ElementType>>()->shared_from_this() : nullptr;
std::shared_ptr<ComputationNode<ElementType>> input0Node = (input0BaseNodePtr != nullptr) ? input0BaseNodePtr->template As<ComputationNode<ElementType>>()->shared_from_this() : nullptr;
std::shared_ptr<ComputationNode<ElementType>> input1Node;
if (functionInputs.size() > 1)
{
auto input1BaseNodePtr = GetNode(functionInputs[1], network, builder, variableToNodeMap, isVariableRootMap);
input1Node = (input1BaseNodePtr != nullptr) ? input1BaseNodePtr->As<ComputationNode<ElementType>>()->shared_from_this() : nullptr;
input1Node = (input1BaseNodePtr != nullptr) ? input1BaseNodePtr->template As<ComputationNode<ElementType>>()->shared_from_this() : nullptr;
}
PrimitiveOpType op = primitiveFunction->OpType();
@ -303,17 +303,17 @@ namespace CNTK
template <typename ElementType>
/*static*/ std::pair<std::shared_ptr<const Matrix<ElementType>>, MBLayoutPtr> CompositeFunction::GetCNTKImplMatrixAndMBLayoutFromValueObject(Variable var, const ValuePtr& value)
{
if (var.DataType() != value->Data()->DataType())
LogicError("The Variable's DataType %s does not match the corresponding Value's DataType %s", DataTypeName(var.DataType()), DataTypeName(value->Data()->DataType()));
if (var.GetDataType() != value->Data()->GetDataType())
LogicError("The Variable's DataType %s does not match the corresponding Value's DataType %s", DataTypeName(var.GetDataType()), DataTypeName(value->Data()->GetDataType()));
if (GetDataType<ElementType>() != value->Data()->DataType())
LogicError("The specified ElementType %s does not match the DataType %s", typeid(ElementType).name(), DataTypeName(value->Data()->DataType()));
if (AsDataType<ElementType>() != value->Data()->GetDataType())
LogicError("The specified ElementType %s does not match the DataType %s", typeid(ElementType).name(), DataTypeName(value->Data()->GetDataType()));
// TODO: Is supplying dense data for an Input variable tagged as sparse, a fatal error?
if (var.IsSparseInput() && !value->Data()->IsSparse())
InvalidArgument("Dense input data supplied for a sparse input Variable");
if (var.IsSparseInput() && (value->Data()->StorageFormat() != StorageFormat::SparseCSC))
if (var.IsSparseInput() && (value->Data()->GetStorageFormat() != StorageFormat::SparseCSC))
InvalidArgument("Sparse Input data must be in SparseCSC format");
if (value->Data()->Shape().NumAxes() == var.Shape().NumAxes())
@ -397,7 +397,7 @@ namespace CNTK
layout->GetNumCols(),
AsCNTKImplDeviceId(value->Data()->Device()),
value->Data()->IsSparse() ? MatrixType::SPARSE : MatrixType::DENSE,
AsCNTKMatrixFormat(value->Data()->StorageFormat()));
AsCNTKMatrixFormat(value->Data()->GetStorageFormat()));
std::vector<size_t> sequencesShorterThanLongestSequence;
for (size_t i = 0; i < numSequences; ++i)
@ -427,8 +427,8 @@ namespace CNTK
if (var.DynamicAxes().size() > 1)
LogicError("More than one dynamic axis for a variable is currently unsupported");
if (GetDataType<ElementType>() != var.DataType())
LogicError("The specified ElementType %s does not match the DataType %s", typeid(ElementType).name(), DataTypeName(var.DataType()));
if (AsDataType<ElementType>() != var.GetDataType())
LogicError("The specified ElementType %s does not match the DataType %s", typeid(ElementType).name(), DataTypeName(var.GetDataType()));
if ((layout != nullptr) && (matrix.GetNumRows() != var.Shape().TotalSize()))
LogicError("Unexpected matrix layout: The number of rows in the matrix does not match the sample size of the Variable");
@ -442,7 +442,7 @@ namespace CNTK
{
// Just create a view over the existing matrix itself
auto tensorView = new TensorView<ElementType>(std::make_shared<Matrix<ElementType>>(matrix.AsReference()), AsTensorShape(valueDataShape));
auto data = NDArrayViewPtr(new NDArrayView(GetDataType<ElementType>(), AsDeviceDescriptor(matrix.GetDeviceId()), AsStorageFormat(matrix.GetFormat()), valueDataShape, true, tensorView), [](_ReferenceCounter* ptr) { delete ptr; });
auto data = NDArrayViewPtr(new NDArrayView(AsDataType<ElementType>(), AsDeviceDescriptor(matrix.GetDeviceId()), AsStorageFormat(matrix.GetFormat()), valueDataShape, true, tensorView), [](_ReferenceCounter* ptr) { delete ptr; });
return ValuePtr(new Value(data), [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -502,7 +502,7 @@ namespace CNTK
}
auto tensorView = new TensorView<ElementType>(shuffledMatrixData, AsTensorShape(valueDataShape));
auto data = NDArrayViewPtr(new NDArrayView(GetDataType<ElementType>(), AsDeviceDescriptor(matrix.GetDeviceId()), StorageFormat::Dense, valueDataShape, true, tensorView), [](_ReferenceCounter* ptr) { delete ptr; });
auto data = NDArrayViewPtr(new NDArrayView(AsDataType<ElementType>(), AsDeviceDescriptor(matrix.GetDeviceId()), StorageFormat::Dense, valueDataShape, true, tensorView), [](_ReferenceCounter* ptr) { delete ptr; });
return ValuePtr(new Value(data, mask), [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -522,7 +522,7 @@ namespace CNTK
ValuePtr argumentValue = arguments[*iter];
MBLayoutPtr layout;
switch (argumentValue->Data()->DataType())
switch (argumentValue->Data()->GetDataType())
{
case DataType::Float:
{
@ -547,7 +547,7 @@ namespace CNTK
break;
}
default:
LogicError("Unsupported DataType %s", DataTypeName(argumentValue->Data()->DataType()));
LogicError("Unsupported DataType %s", DataTypeName(argumentValue->Data()->GetDataType()));
break;
}
@ -573,7 +573,7 @@ namespace CNTK
ValuePtr gradientValue = iter->second;
MBLayoutPtr layout;
switch (gradientValue->Data()->DataType())
switch (gradientValue->Data()->GetDataType())
{
case DataType::Float:
{
@ -594,7 +594,7 @@ namespace CNTK
break;
}
default:
LogicError("Unsupported DataType %s", DataTypeName(gradientValue->Data()->DataType()));
LogicError("Unsupported DataType %s", DataTypeName(gradientValue->Data()->GetDataType()));
break;
}
}
@ -635,14 +635,14 @@ namespace CNTK
InvalidArgument("The shape %s of the specified Value object for output does not match the actual output shape %s", AsString(outputValuePtr->Data()->Shape()).c_str(), AsString(outputShape).c_str());
}
switch (iter->first.DataType())
switch (iter->first.GetDataType())
{
case DataType::Float:
{
auto nodeValue = GetValueObjectFromCNTKImplMatrixAndMBLayout<float>(iter->first, computationNodePtr->As<ComputationNode<float>>()->Value(), computationNodePtr->GetMBLayout());
if (outputValuePtr == nullptr)
{
auto data = NDArrayViewPtr(new NDArrayView(iter->first.DataType(), outputShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto data = NDArrayViewPtr(new NDArrayView(iter->first.GetDataType(), outputShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto mask = (nodeValue->Mask() != nullptr) ? NDMaskPtr(new NDMask(nodeValue->Mask()->Shape(), nodeValue->Mask()->Device()), [](_ReferenceCounter* ptr) { delete ptr; }) : nullptr;
outputValuePtr = ValuePtr(new Value(data, mask), [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -654,7 +654,7 @@ namespace CNTK
auto nodeValue = GetValueObjectFromCNTKImplMatrixAndMBLayout<double>(iter->first, computationNodePtr->As<ComputationNode<double>>()->Value(), computationNodePtr->GetMBLayout());
if (outputValuePtr == nullptr)
{
auto data = NDArrayViewPtr(new NDArrayView(iter->first.DataType(), outputShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto data = NDArrayViewPtr(new NDArrayView(iter->first.GetDataType(), outputShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto mask = (nodeValue->Mask() != nullptr) ? NDMaskPtr(new NDMask(nodeValue->Mask()->Shape(), nodeValue->Mask()->Device()), [](_ReferenceCounter* ptr) { delete ptr; }) : nullptr;
outputValuePtr = ValuePtr(new Value(data, mask), [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -662,7 +662,7 @@ namespace CNTK
break;
}
default:
LogicError("Unsupported DataType %s", DataTypeName(iter->first.DataType()));
LogicError("Unsupported DataType %s", DataTypeName(iter->first.GetDataType()));
break;
}
@ -698,14 +698,14 @@ namespace CNTK
if (!computationNodePtr->NeedsGradient())
LogicError("Backpropagated gradient value cannot be read from a ComputationNode that has NeedsGradient set to false");
switch (iter->first.DataType())
switch (iter->first.GetDataType())
{
case DataType::Float:
{
auto nodeValue = GetValueObjectFromCNTKImplMatrixAndMBLayout<float>(iter->first, computationNodePtr->As<ComputationNode<float>>()->Gradient(), computationNodePtr->GetMBLayout());
if (gradientValuePtr == nullptr)
{
auto data = NDArrayViewPtr(new NDArrayView(iter->first.DataType(), gradientShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto data = NDArrayViewPtr(new NDArrayView(iter->first.GetDataType(), gradientShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto mask = NDMaskPtr((nodeValue->Mask() != nullptr) ? new NDMask(nodeValue->Mask()->Shape(), nodeValue->Mask()->Device()) : nullptr, [](_ReferenceCounter* ptr) { delete ptr; });
gradientValuePtr = ValuePtr(new Value(data, mask), [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -717,7 +717,7 @@ namespace CNTK
auto nodeValue = GetValueObjectFromCNTKImplMatrixAndMBLayout<double>(iter->first, computationNodePtr->As<ComputationNode<double>>()->Gradient(), computationNodePtr->GetMBLayout());
if (gradientValuePtr == nullptr)
{
auto data = NDArrayViewPtr(new NDArrayView(iter->first.DataType(), gradientShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto data = NDArrayViewPtr(new NDArrayView(iter->first.GetDataType(), gradientShape, AsDeviceDescriptor(computationNodePtr->ValuePtr()->GetDeviceId())), [](_ReferenceCounter* ptr) { delete ptr; });
auto mask = NDMaskPtr((nodeValue->Mask() != nullptr) ? new NDMask(nodeValue->Mask()->Shape(), nodeValue->Mask()->Device()) : nullptr, [](_ReferenceCounter* ptr) { delete ptr; });
gradientValuePtr = ValuePtr(new Value(data, mask), [](_ReferenceCounter* ptr) { delete ptr; });
@ -726,7 +726,7 @@ namespace CNTK
break;
}
default:
LogicError("Unsupported DataType %s", DataTypeName(iter->first.DataType()));
LogicError("Unsupported DataType %s", DataTypeName(iter->first.GetDataType()));
break;
}
@ -741,7 +741,7 @@ namespace CNTK
{
// TODO: How about zero argument functions?
// TODO: We need a better way to determine the ElementType for the network
auto dataType = arguments.m_map->begin()->second->Data()->DataType();
auto dataType = arguments.m_map->begin()->second->Data()->GetDataType();
if (dataType == DataType::Float)
GetComputationNetwork<float>(computeDevice, outputsToRetainBackwardStateFor);
else

4
Source/CNTKv2LibraryDll/Function.h
Просмотреть файл

@ -150,7 +150,7 @@ namespace CNTK
static NDShape ReductionOpOutputShape(PrimitiveOpType op, const NDShape& operandShape, const std::vector<size_t>& reductionAxes)
{
if (reductionAxes.size() > operandShape.NumAxes())
RuntimeError("The number of reduction axes %d exceeds the number of axes in the operand shape %s of the reduction operation %s", reductionAxes.size(), AsString(operandShape).c_str(), PrimitiveOpTypeName(op));
RuntimeError("The number of reduction axes %d exceeds the number of axes in the operand shape %s of the reduction operation %s", (int)reductionAxes.size(), AsString(operandShape).c_str(), PrimitiveOpTypeName(op));
size_t numOutputAxes = operandShape.NumAxes() - reductionAxes.size();
std::vector<size_t> outputDims(numOutputAxes);
@ -171,7 +171,7 @@ namespace CNTK
std::vector<Variable> outputs;
// TODO: We are just using the input[0]'s DataType as output node's DataType. This is not always correct
DataType outputDataType = inputs[0].DataType();
DataType outputDataType = inputs[0].GetDataType();
// We currently require that the inputs' dynamic axes if any match
std::vector<Axis> outputDynamicAxes = inputs[0].DynamicAxes();

2
Source/CNTKv2LibraryDll/LibraryTests/Common.h
Просмотреть файл

@ -15,6 +15,6 @@ inline void FloatingPointVectorCompare(const std::vector<ElementType>& first, co
ElementType rightVal = second[i];
ElementType allowedTolerance = (std::max<ElementType>)((ElementType)absoluteTolerance, ((ElementType)relativeTolerance) * leftVal);
if (std::abs(leftVal - rightVal) > allowedTolerance)
throw std::exception(message);
throw std::runtime_error(message);
}
}

22
Source/CNTKv2LibraryDll/LibraryTests/FeedForwardTests.cpp
Просмотреть файл

@ -50,13 +50,13 @@ void TestFeedForwardNetworkCreation(const DeviceDescriptor& device)
// Now test the structure
if (ffNet->Parameters().size() != ((numHiddenLayers * 2) + 1))
throw std::exception("TestFeedForwardNetworkCreation: Function does not have expected Parameter count");
throw std::runtime_error("TestFeedForwardNetworkCreation: Function does not have expected Parameter count");
if (ffNet->Arguments().size() != 2)
throw std::exception("TestFeedForwardNetworkCreation: Function does not have expected Argument count");
throw std::runtime_error("TestFeedForwardNetworkCreation: Function does not have expected Argument count");
if (ffNet->Outputs().size() != 3)
throw std::exception("TestFeedForwardNetworkCreation: Function does not have expected Output count");
throw std::runtime_error("TestFeedForwardNetworkCreation: Function does not have expected Output count");
// Run Forward and backward a few times
size_t iterationCount = 4;
@ -109,7 +109,7 @@ void TestTimesAndPlus(size_t inputDim,
Parameter timesParam(new NDArrayView((ElementType)0.5, { outputDim, inputDim }, device));
Parameter plusParam(new NDArrayView((ElementType)1.2, { outputDim }, device));
Variable inputVar({ inputDim }, GetDataType<ElementType>(), L"input");
Variable inputVar({ inputDim }, AsDataType<ElementType>(), L"input");
auto timesAndPlusFunc = Plus(plusParam, Times(timesParam, inputVar));
srand(seed);
@ -131,7 +131,7 @@ void TestTimesAndPlus(size_t inputDim,
if (outputAllocationDevice.Type() == DeviceType::CPU)
outputValue = new Value(new NDArrayView(outputShape, outputData.data(), outputData.size(), outputAllocationDevice, false));
else
outputValue = new Value(new NDArrayView(GetDataType<ElementType>(), outputShape, outputAllocationDevice));
outputValue = new Value(new NDArrayView(AsDataType<ElementType>(), outputShape, outputAllocationDevice));
}
std::unordered_map<Variable, ValuePtr> outputs = { { timesAndPlusFunc->Output(), outputValue } };
@ -148,7 +148,7 @@ void TestTimesAndPlus(size_t inputDim,
else
{
NDArrayViewPtr cpuArrayView = new NDArrayView(outputShape, rootGradientsData.data(), rootGradientsData.size(), DeviceDescriptor::CPUDevice(), true);
NDArrayViewPtr gpuArrayView = new NDArrayView(GetDataType<ElementType>(), outputShape, device);
NDArrayViewPtr gpuArrayView = new NDArrayView(AsDataType<ElementType>(), outputShape, device);
gpuArrayView->CopyFrom(*cpuArrayView);
rootGradientValue = new Value(gpuArrayView);
}
@ -166,8 +166,8 @@ void TestTimesAndPlus(size_t inputDim,
}
else
{
plusParameterGradientValue = new Value(new NDArrayView(GetDataType<ElementType>(), plusParam.Shape(), outputAllocationDevice));
timesParameterGradientValue = new Value(new NDArrayView(GetDataType<ElementType>(), timesParam.Shape(), outputAllocationDevice));
plusParameterGradientValue = new Value(new NDArrayView(AsDataType<ElementType>(), plusParam.Shape(), outputAllocationDevice));
timesParameterGradientValue = new Value(new NDArrayView(AsDataType<ElementType>(), timesParam.Shape(), outputAllocationDevice));
}
}
@ -203,12 +203,12 @@ void TestTimesAndPlus(size_t inputDim,
// Verify backward prop results
if (device.Type() != DeviceType::CPU)
{
NDArrayViewPtr cpuArrayView = new NDArrayView(GetDataType<ElementType>(), plusParam.Shape(), DeviceDescriptor::CPUDevice());
NDArrayViewPtr cpuArrayView = new NDArrayView(AsDataType<ElementType>(), plusParam.Shape(), DeviceDescriptor::CPUDevice());
cpuArrayView->CopyFrom(*plusParameterGradientValue->Data());
const ElementType* cpuArrayViewBuffer = cpuArrayView->DataBuffer<ElementType>();
memcpy(plusParameterGradientData.data(), cpuArrayViewBuffer, plusParam.Shape().TotalSize() * sizeof(ElementType));
cpuArrayView = new NDArrayView(GetDataType<ElementType>(), timesParam.Shape(), DeviceDescriptor::CPUDevice());
cpuArrayView = new NDArrayView(AsDataType<ElementType>(), timesParam.Shape(), DeviceDescriptor::CPUDevice());
cpuArrayView->CopyFrom(*timesParameterGradientValue->Data());
cpuArrayViewBuffer = cpuArrayView->DataBuffer<ElementType>();
memcpy(timesParameterGradientData.data(), cpuArrayViewBuffer, timesParam.Shape().TotalSize() * sizeof(ElementType));
@ -216,7 +216,7 @@ void TestTimesAndPlus(size_t inputDim,
for (size_t i = 0; i < outputDim; ++i)
if (plusParameterGradientData[i] != numSamples)
throw std::exception("TestTimesAndPlus: Backprop prop results do not match expected results for Plus params gradients");
throw std::runtime_error("TestTimesAndPlus: Backprop prop results do not match expected results for Plus params gradients");
std::vector<ElementType> expectedTimesParamsGradientValues(timesParam.Shape().TotalSize());
for (size_t i = 0; i < inputDim; ++i)

32
Source/CNTKv2LibraryDll/LibraryTests/NDArrayViewTests.cpp
Просмотреть файл

@ -18,7 +18,7 @@ void TestNDArrayView(size_t numAxes, const DeviceDescriptor& device)
std::array<ElementType, 1> arrayData = { 3 };
auto arrayDataView = new NDArrayView({}, arrayData);
if (arrayDataView->DataBuffer<ElementType>() != arrayData.data())
throw std::exception("The DataBuffer of the NDArrayView does not match the original buffer it was created over");
throw std::runtime_error("The DataBuffer of the NDArrayView does not match the original buffer it was created over");
std::vector<ElementType> data(viewShape.TotalSize());
ElementType scale = 19.0;
@ -28,19 +28,19 @@ void TestNDArrayView(size_t numAxes, const DeviceDescriptor& device)
auto cpuDataView = new NDArrayView(viewShape, data);
if (cpuDataView->DataBuffer<ElementType>() != data.data())
throw std::exception("The DataBuffer of the NDArrayView does not match the original buffer it was created over");
throw std::runtime_error("The DataBuffer of the NDArrayView does not match the original buffer it was created over");
NDArrayViewPtr dataView;
if ((device.Type() == DeviceType::CPU))
dataView = cpuDataView;
else
{
dataView = new NDArrayView(GetDataType<ElementType>(), viewShape, device);
dataView = new NDArrayView(AsDataType<ElementType>(), viewShape, device);
dataView->CopyFrom(*cpuDataView);
}
if (dataView->Device() != device)
throw std::exception("Device of NDArrayView does not match 'device' it was created on");
throw std::runtime_error("Device of NDArrayView does not match 'device' it was created on");
// Test clone
auto clonedView = dataView->DeepClone(false);
@ -50,13 +50,13 @@ void TestNDArrayView(size_t numAxes, const DeviceDescriptor& device)
if ((device.Type() == DeviceType::CPU))
{
if (dataView->DataBuffer<ElementType>() != data.data())
throw std::exception("The DataBuffer of the NDArrayView does not match the original buffer it was created over");
throw std::runtime_error("The DataBuffer of the NDArrayView does not match the original buffer it was created over");
first = clonedView->WritableDataBuffer<ElementType>();
}
else
{
temp1CpuDataView = new NDArrayView(GetDataType<ElementType>(), viewShape, DeviceDescriptor::CPUDevice());
temp1CpuDataView = new NDArrayView(AsDataType<ElementType>(), viewShape, DeviceDescriptor::CPUDevice());
temp1CpuDataView->CopyFrom(*clonedView);
first = temp1CpuDataView->WritableDataBuffer<ElementType>();
@ -65,7 +65,7 @@ void TestNDArrayView(size_t numAxes, const DeviceDescriptor& device)
for (size_t i = 0; i < viewShape.TotalSize(); ++i)
{
if (first[i] != second[i])
throw std::exception("The contents of the clone do not match expected");
throw std::runtime_error("The contents of the clone do not match expected");
}
first[0] += 1;
@ -79,28 +79,28 @@ void TestNDArrayView(size_t numAxes, const DeviceDescriptor& device)
}
else
{
temp1CpuDataView = new NDArrayView(GetDataType<ElementType>(), viewShape, DeviceDescriptor::CPUDevice());
temp1CpuDataView = new NDArrayView(AsDataType<ElementType>(), viewShape, DeviceDescriptor::CPUDevice());
temp1CpuDataView->CopyFrom(*clonedView);
first = temp1CpuDataView->WritableDataBuffer<ElementType>();
temp2CpuDataView = new NDArrayView(GetDataType<ElementType>(), viewShape, DeviceDescriptor::CPUDevice());
temp2CpuDataView = new NDArrayView(AsDataType<ElementType>(), viewShape, DeviceDescriptor::CPUDevice());
temp2CpuDataView->CopyFrom(*dataView);
second = temp2CpuDataView->DataBuffer<ElementType>();
}
if (first[0] != (second[0] + 1))
throw std::exception("The clonedView's contents do not match expected");
throw std::runtime_error("The clonedView's contents do not match expected");
// Test alias
auto aliasView = clonedView->Alias(true);
const ElementType* aliasViewBuffer = aliasView->DataBuffer<ElementType>();
const ElementType* clonedDataBuffer = clonedView->DataBuffer<ElementType>();
if (aliasViewBuffer != clonedDataBuffer)
throw std::exception("The buffers underlying the alias view and the view it is an alias of are different!");
throw std::runtime_error("The buffers underlying the alias view and the view it is an alias of are different!");
clonedView->CopyFrom(*dataView);
if (aliasViewBuffer != clonedDataBuffer)
throw std::exception("The buffers underlying the alias view and the view it is an alias of are different!");
throw std::runtime_error("The buffers underlying the alias view and the view it is an alias of are different!");
// Test readonliness
auto verifyException = [](const std::function<void()>& functionToTest) {
@ -115,7 +115,7 @@ void TestNDArrayView(size_t numAxes, const DeviceDescriptor& device)
}
if (!error)
throw std::exception("Was incorrectly able to get a writable buffer pointer from a readonly view");
throw std::runtime_error("Was incorrectly able to get a writable buffer pointer from a readonly view");
};
// Should not be able to get the WritableDataBuffer for a read-only view
@ -184,14 +184,14 @@ void TestSparseCSCArrayView(size_t numAxes, const DeviceDescriptor& device)
NDArrayView denseCPUTensor(viewShape, copiedDenseData.data(), copiedDenseData.size(), DeviceDescriptor::CPUDevice());
denseCPUTensor.CopyFrom(sparseCSCArrayView);
if (copiedDenseData != referenceDenseData)
throw std::exception("The contents of the dense vector that the sparse NDArrayView is copied into do not match the expected values");
throw std::runtime_error("The contents of the dense vector that the sparse NDArrayView is copied into do not match the expected values");
NDArrayView emptySparseCSCArrayView(GetDataType<ElementType>(), StorageFormat::SparseCSC, viewShape, device);
NDArrayView emptySparseCSCArrayView(AsDataType<ElementType>(), StorageFormat::SparseCSC, viewShape, device);
emptySparseCSCArrayView.CopyFrom(denseCPUTensor);
NDArrayView newDenseCPUTensor(viewShape, copiedDenseData.data(), copiedDenseData.size(), DeviceDescriptor::CPUDevice());
newDenseCPUTensor.CopyFrom(emptySparseCSCArrayView);
if (copiedDenseData != referenceDenseData)
throw std::exception("The contents of the dense vector that the sparse NDArrayView is copied into do not match the expected values");
throw std::runtime_error("The contents of the dense vector that the sparse NDArrayView is copied into do not match the expected values");
}
void NDArrayViewTests()

24
Source/CNTKv2LibraryDll/LibraryTests/RecurrentFunctionTests.cpp
Просмотреть файл

@ -145,10 +145,10 @@ void TestRecurrentNetworkCreation(const DeviceDescriptor& device)
const size_t hiddenDim = 512;
const size_t numOutputClasses = 9304;
Variable features({ inputDim }, GetDataType<ElementType>(), L"Features");
Variable features({ inputDim }, AsDataType<ElementType>(), L"Features");
auto classifierOutputFunction = LSTMNet<ElementType>(features, cellDim, hiddenDim, numOutputClasses, numLSTMLayers, device);
Variable labelsVar = Variable({ numOutputClasses }, GetDataType<ElementType>(), L"Labels");
Variable labelsVar = Variable({ numOutputClasses }, AsDataType<ElementType>(), L"Labels");
auto trainingLossFunction = CrossEntropyWithSoftmax(classifierOutputFunction, labelsVar, L"lossFunction");
auto predictionFunction = PredictionError(classifierOutputFunction, labelsVar, L"predictionError");
@ -156,13 +156,13 @@ void TestRecurrentNetworkCreation(const DeviceDescriptor& device)
// Now test the structure
if (LSTMClassifier->Arguments().size() != 2)
throw std::exception("TestFeedForwardNetworkCreation: Function does not have expected Argument count");
throw std::runtime_error("TestFeedForwardNetworkCreation: Function does not have expected Argument count");
if (LSTMClassifier->Outputs().size() != 3)
throw std::exception("TestFeedForwardNetworkCreation: Function does not have expected Output count");
throw std::runtime_error("TestFeedForwardNetworkCreation: Function does not have expected Output count");
if (LSTMClassifier->Parameters().size() != ((numLSTMLayers * 28) + 3))
throw std::exception("TestFeedForwardNetworkCreation: Function does not have expected Parameter count");
throw std::runtime_error("TestFeedForwardNetworkCreation: Function does not have expected Parameter count");
// Run Forward and backward a few times
size_t iterationCount = 3;
@ -181,7 +181,7 @@ void TestRecurrentNetworkCreation(const DeviceDescriptor& device)
maxActualSequenceLength = sequenceLengths[i];
}
std::vector<const std::vector<ElementType>> inputSequences;
std::vector<std::vector<ElementType>> inputSequences;
for (size_t i = 0; i < numSequences; ++i)
{
std::vector<ElementType> currentSequence(inputDim * sequenceLengths[i]);
@ -193,7 +193,7 @@ void TestRecurrentNetworkCreation(const DeviceDescriptor& device)
ValuePtr inputValue = Value::Create({ inputDim }, inputSequences, device, true);
std::vector<const std::vector<ElementType>> labelsData;
std::vector<std::vector<ElementType>> labelsData;
for (size_t i = 0; i < numSequences; ++i)
{
std::vector<ElementType> currentSequence(numOutputClasses * sequenceLengths[i]);
@ -235,12 +235,12 @@ void TestSimpleRecurrence(size_t inputDim,
unsigned int seed = 1)
{
if (useOneHotSparseInputs && !useSparseInputs)
throw std::exception("useOneHotSparseInputs option can only be true when useSparseInputs is true");
throw std::runtime_error("useOneHotSparseInputs option can only be true when useSparseInputs is true");
Parameter timesParam(new NDArrayView((ElementType)0.5, { outputDim, inputDim }, device));
Parameter plusParam(new NDArrayView((ElementType)0.1, { outputDim }, device));
Variable inputVar({ inputDim }, useSparseInputs, GetDataType<ElementType>(), true, L"input");
Variable inputVar({ inputDim }, useSparseInputs, AsDataType<ElementType>(), true, L"input");
auto placeholder = Placeholder({ outputDim });
auto plusOutput = Plus(plusParam, Plus(placeholder, Times(timesParam, inputVar)));
@ -274,7 +274,7 @@ void TestSimpleRecurrence(size_t inputDim,
std::vector<ElementType> inputData(inputDim * totalNumInputSamples, useSparseInputs ? 0 : std::numeric_limits<ElementType>::quiet_NaN());
if (useOneHotSparseInputs)
{
std::vector<const std::vector<size_t>> oneHotSequences;
std::vector<std::vector<size_t>> oneHotSequences;
for (size_t i = 0; i < numSequences; ++i)
{
std::vector<size_t> currentSequence(sequenceLengths[i]);
@ -314,7 +314,7 @@ void TestSimpleRecurrence(size_t inputDim,
NDArrayViewPtr inputValueData = new NDArrayView(inputShape, inputData.data(), inputData.size(), DeviceDescriptor::CPUDevice(), true);
if (useSparseInputs)
{
NDArrayViewPtr sparseInputValueData = new NDArrayView(GetDataType<ElementType>(), StorageFormat::SparseCSC, inputShape, DeviceDescriptor::CPUDevice());
NDArrayViewPtr sparseInputValueData = new NDArrayView(AsDataType<ElementType>(), StorageFormat::SparseCSC, inputShape, DeviceDescriptor::CPUDevice());
sparseInputValueData->CopyFrom(*inputValueData);
inputValueData = sparseInputValueData->Alias(true);
}
@ -423,7 +423,7 @@ void TestSimpleRecurrence(size_t inputDim,
for (size_t k = 0; k < plusParam.Shape().TotalSize(); ++k)
if (plusParameterGradientData[k] != expectedPlusParameterGradientValue)
throw std::exception("TestSimpleRecurrence: Backprop prop results do not match expected results for Plus params gradients");
throw std::runtime_error("TestSimpleRecurrence: Backprop prop results do not match expected results for Plus params gradients");
std::vector<ElementType> expectedTimesParamsGradientValues(timesParam.Shape().TotalSize(), 0);
for (size_t i = 0; i < numSequences; ++i)

4
Source/CNTKv2LibraryDll/LibraryTests/TensorTests.cpp
Просмотреть файл

@ -20,8 +20,8 @@ void TestTensorPlus(size_t numAxesLeftOperand, size_t numAxesRightOperand, const
for (size_t i = std::min(numAxesLeftOperand, numAxesRightOperand); i < numAxesRightOperand; ++i)
rightInputShape[i] = (rand() % maxDimSize) + 1;
Variable leftInputVar(leftInputShape, GetDataType<ElementType>(), L"leftInput");
Variable rightInputVar(rightInputShape, GetDataType<ElementType>(), L"rightInput");
Variable leftInputVar(leftInputShape, AsDataType<ElementType>(), L"leftInput");
Variable rightInputVar(rightInputShape, AsDataType<ElementType>(), L"rightInput");
auto plusFunc = Plus(leftInputVar, rightInputVar);

19
Source/CNTKv2LibraryDll/NDArrayView.cpp
Просмотреть файл

@ -88,7 +88,7 @@ namespace CNTK
template <typename ElementType>
NDArrayView::NDArrayView(const NDShape& viewShape, const SparseIndexType* colStarts, const SparseIndexType* rowIndices, const ElementType* nonZeroValues, size_t numNonZeroValues, const DeviceDescriptor& device, bool readOnly/* = false*/)
: NDArrayView(GetDataType<ElementType>(), device, StorageFormat::SparseCSC, viewShape, false, AllocateTensorView<ElementType>(viewShape, StorageFormat::SparseCSC, device))
: NDArrayView(AsDataType<ElementType>(), device, StorageFormat::SparseCSC, viewShape, false, AllocateTensorView<ElementType>(viewShape, StorageFormat::SparseCSC, device))
{
if ((colStarts == nullptr) || (rowIndices == nullptr) || (nonZeroValues == nullptr) || (numNonZeroValues == 0) || (numNonZeroValues > viewShape.TotalSize()))
InvalidArgument("Invalid sparse CSC format initial data specified for NDArrayView construction");
@ -194,7 +194,7 @@ namespace CNTK
template <typename ElementType>
const TensorView<ElementType>* NDArrayView::GetTensorView() const
{
if (GetDataType<ElementType>() != m_dataType)
if (AsDataType<ElementType>() != m_dataType)
LogicError("NDArrayView::GetWritableTensorView: The specified ElementType %s does not match the DataType %s", typeid(ElementType).name(), DataTypeName(m_dataType));
return (const TensorView<ElementType>*)(m_tensorView);
@ -211,7 +211,7 @@ namespace CNTK
NDArrayViewPtr NDArrayView::DeepClone(bool readOnly/* = false*/) const
{
NDArrayViewPtr newView(new NDArrayView(this->DataType(), this->StorageFormat(), this->Shape(), this->Device()), [](_ReferenceCounter* ptr) { delete ptr; });
NDArrayViewPtr newView(new NDArrayView(this->GetDataType(), this->GetStorageFormat(), this->Shape(), this->Device()), [](_ReferenceCounter* ptr) { delete ptr; });
switch (m_dataType)
{
case DataType::Float:
@ -285,7 +285,7 @@ namespace CNTK
break;
}
auto aliasView = new NDArrayView(DataType(), Device(), StorageFormat(), Shape(), IsReadOnly() || readOnly, tensorView);;
auto aliasView = new NDArrayView(GetDataType(), Device(), GetStorageFormat(), Shape(), IsReadOnly() || readOnly, tensorView);;
return NDArrayViewPtr(aliasView, [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -303,7 +303,7 @@ namespace CNTK
template <typename ElementType>
const ElementType* NDArrayView::DataBuffer() const
{
if (GetDataType<ElementType>() != m_dataType)
if (AsDataType<ElementType>() != m_dataType)
LogicError("The specified ElementType %s does not match the DataType %s", typeid(ElementType).name(), DataTypeName(m_dataType));
if (IsSparse())
@ -322,7 +322,7 @@ namespace CNTK
auto randomUniformMatrix = std::make_shared<Matrix<ElementType>>(Matrix<ElementType>::RandomUniform(matrixDims.first, matrixDims.second, AsCNTKImplDeviceId(device), (ElementType)rangeStart, (ElementType)rangeEnd, seed));
auto tensorView = new TensorView<ElementType>(randomUniformMatrix, AsTensorShape(shape));
auto view = new NDArrayView(GetDataType<ElementType>(), device, StorageFormat::Dense, shape, false, tensorView);
auto view = new NDArrayView(AsDataType<ElementType>(), device, StorageFormat::Dense, shape, false, tensorView);
return NDArrayViewPtr(view, [](_ReferenceCounter* ptr) { delete ptr; });
}
@ -336,8 +336,11 @@ namespace CNTK
template CNTK_API float* NDArrayView::WritableDataBuffer<float>();
template CNTK_API double* NDArrayView::WritableDataBuffer<double>();
template CNTK_API NDArrayView::NDArrayView(const NDShape& viewShape, const SparseIndexType* colStarts, const SparseIndexType* rowIndices, const float* nonZeroValues, size_t numNonZeroValues, const DeviceDescriptor& device, bool readOnly/* = false*/);
template CNTK_API NDArrayView::NDArrayView(const NDShape& viewShape, const SparseIndexType* colStarts, const SparseIndexType* rowIndices, const double* nonZeroValues, size_t numNonZeroValues, const DeviceDescriptor& device, bool readOnly/* = false*/);
template std::shared_ptr<const Matrix<float>> NDArrayView::GetMatrix(size_t rowColSplitPoint/* = AutoSelectRowColSplitPoint*/) const;
template std::shared_ptr<const Matrix<double>> NDArrayView::GetMatrix(size_t rowColSplitPoint/* = AutoSelectRowColSplitPoint*/) const;
template std::shared_ptr<Matrix<float>> NDArrayView::GetWritableMatrix(size_t rowColSplitPoint/* = AutoSelectRowColSplitPoint*/);
template std::shared_ptr<Matrix<double>> NDArrayView::GetWritableMatrix(size_t rowColSplitPoint/* = AutoSelectRowColSplitPoint*/);
template CNTK_API NDArrayView::NDArrayView(const NDShape& viewShape, const SparseIndexType* colStarts, const SparseIndexType* rowIndices, const float* nonZeroValues, size_t numNonZeroValues, const DeviceDescriptor& device, bool readOnly/* = false*/);
template CNTK_API NDArrayView::NDArrayView(const NDShape& viewShape, const SparseIndexType* colStarts, const SparseIndexType* rowIndices, const double* nonZeroValues, size_t numNonZeroValues, const DeviceDescriptor& device, bool readOnly/* = false*/);

6
Source/CNTKv2LibraryDll/Utils.cpp
Просмотреть файл

@ -122,8 +122,11 @@ namespace CNTK
// Explicit template instantiations
template class _SimpleVector<Variable>;
template class _SimpleVector<size_t>;
template class _SimpleVector<Axis>;
template class _SimpleVector<FunctionPtr>;
template bool operator==(const _SimpleVector<size_t>& first, const _SimpleVector<size_t>& second);
#pragma endregion _SimpleVector
#pragma region _SimpleSet
@ -215,9 +218,11 @@ namespace CNTK
// Explicit template instantiations
template class _SimpleSet<FunctionPtr>;
template class _SimpleSet<Variable>;
template class _SimpleSet<Placeholder>;
template class _SimpleSet<const Function*>;
template bool operator==(const _SimpleSet<Variable>& first, const _SimpleSet<Variable>& second);
template bool operator==(const _SimpleSet<Placeholder>& first, const _SimpleSet<Placeholder>& second);
#pragma endregion _SimpleSet
@ -314,6 +319,7 @@ namespace CNTK
}
// Explicit template instantiations
template class _SimpleMap<Variable, ValuePtr>;
template class _SimpleMap<Variable, const ValuePtr>;
template class _SimpleMap<Placeholder, Variable>;

2
Source/CNTKv2LibraryDll/Utils.h
Просмотреть файл

@ -75,7 +75,7 @@ namespace CNTK
: m_valueType(GetValueType<T>())
{
static_assert(std::is_same<T, NDShape>::value ||
std::is_same<T, _Internal::_SimpleVector<DictionaryValue>::value,
std::is_same<T, _Internal::_SimpleVector<DictionaryValue>>::value,
"Unsupported ValueType");
AllocateDataPtr(value);

18
Source/CNTKv2LibraryDll/Value.cpp
Просмотреть файл

@ -28,7 +28,7 @@ namespace CNTK
}
template <typename T>
static NDMaskPtr CreateMask(size_t sampleSize, const std::vector<const std::vector<T>>& sequences, const DeviceDescriptor& device)
static NDMaskPtr CreateMask(size_t sampleSize, const std::vector<std::vector<T>>& sequences, const DeviceDescriptor& device)
{
size_t numSequences = sequences.size();
std::vector<size_t> sequenceLengths(numSequences);
@ -58,7 +58,7 @@ namespace CNTK
}
template <typename ElementType>
/*static*/ ValuePtr Value::Create(size_t vocabularySize, const std::vector<const std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly/* = false*/)
/*static*/ ValuePtr Value::Create(size_t vocabularySize, const std::vector<std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly/* = false*/)
{
NDMaskPtr deviceValueMask = CreateMask(1, oneHotSequences, device);
size_t maxSequenceLength = (deviceValueMask == nullptr) ? oneHotSequences[0].size() : deviceValueMask->Shape()[0];
@ -91,7 +91,7 @@ namespace CNTK
}
template <typename ElementType>
/*static*/ ValuePtr Value::Create(const NDShape& sampleShape, const std::vector<const std::vector<ElementType>>& sequences, const DeviceDescriptor& device, bool readOnly/* = false*/)
/*static*/ ValuePtr Value::Create(const NDShape& sampleShape, const std::vector<std::vector<ElementType>>& sequences, const DeviceDescriptor& device, bool readOnly/* = false*/)
{
size_t sampleSize = sampleShape.TotalSize();
NDMaskPtr deviceValueMask = CreateMask(sampleSize, sequences, device);
@ -99,7 +99,7 @@ namespace CNTK
size_t numSequences = sequences.size();
NDShape valueDataShape = sampleShape.AppendShape({ maxSequenceLength, numSequences });
NDArrayViewPtr valueData(new NDArrayView(GetDataType<ElementType>(), valueDataShape, DeviceDescriptor::CPUDevice()), [](_ReferenceCounter* ptr) { delete ptr; });
NDArrayViewPtr valueData(new NDArrayView(AsDataType<ElementType>(), valueDataShape, DeviceDescriptor::CPUDevice()), [](_ReferenceCounter* ptr) { delete ptr; });
ElementType* dataBuffer = valueData->WritableDataBuffer<ElementType>();
for (size_t i = 0; i < numSequences; ++i)
std::copy(sequences[i].data(), sequences[i].data() + sequences[i].size(), dataBuffer + (maxSequenceLength * i * sampleSize));
@ -114,7 +114,7 @@ namespace CNTK
}
else
{
deviceValueData = NDArrayViewPtr(new NDArrayView(GetDataType<ElementType>(), valueDataShape, device), [](_ReferenceCounter* ptr) { delete ptr; });
deviceValueData = NDArrayViewPtr(new NDArrayView(AsDataType<ElementType>(), valueDataShape, device), [](_ReferenceCounter* ptr) { delete ptr; });
deviceValueData->CopyFrom(*valueData);
if (readOnly)
deviceValueData = deviceValueData->Alias(true);
@ -171,8 +171,8 @@ namespace CNTK
}
// Explicit template instantiations
template /*static*/ CNTK_API ValuePtr Value::Create<float>(const NDShape& sampleShape, const std::vector<const std::vector<float>>& sequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<double>(const NDShape& sampleShape, const std::vector<const std::vector<double>>& sequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<float>(size_t vocabSize, const std::vector<const std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<double>(size_t vocabSize, const std::vector<const std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<float>(const NDShape& sampleShape, const std::vector<std::vector<float>>& sequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<double>(const NDShape& sampleShape, const std::vector<std::vector<double>>& sequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<float>(size_t vocabSize, const std::vector<std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
template /*static*/ CNTK_API ValuePtr Value::Create<double>(size_t vocabSize, const std::vector<std::vector<size_t>>& oneHotSequences, const DeviceDescriptor& device, bool readOnly/* = false*/);
}

1
Source/Math/GPUMatrix.cu
Просмотреть файл

@ -4412,6 +4412,7 @@ template void GPUMatrix<char>::SetValue(GPUMatrix<char> const&);
template void GPUMatrix<char>::CopySection(size_t numRows, size_t numCols, char* dst, size_t colStride) const;
template void GPUMatrix<char>::Reshape(const size_t, const size_t);
template GPUMatrix<char>& GPUMatrix<char>::operator*=(char);
template DEVICEID_TYPE GPUMatrix<char>::PrepareDevice(DEVICEID_TYPE deviceId) const;
template GPUMatrix<int>::GPUMatrix(const size_t, const size_t, int, int*, const size_t);
template GPUMatrix<int>::~GPUMatrix();