Tensor¶

class Tensor¶

The base tensor class, manage memory or not.

Tensor is usually constructed with the shape info:

auto* a = new dragon::Tensor(std::vector<int64_t>({2, 3}));
auto* b = dragon::Tensor().Reshape({2, 3}); // Equivalent

To allocate the data, type meta and device context are also required:

auto meta = dragon::TypeMeta::Make<float>();
auto* raw_data = a->raw_mutable_data<dragon::CPUContext>(meta);
auto* data = b->mutable_data<float, dragon::CPUContext>();

Memory will be reset if required bytes is larger than capacity:

std::cout << a->nbytes() << " " << a->capacity() << std::endl; // 24, 24
std::cout << a->Reshape({2, 4})->size() << std::endl; // 8
std::cout << a->nbytes() << " " << a->capacity() << std::endl; // 32, 0
a->mutable_data<float, dragon::CPUContext>();
a->Reshape({2, 3});
std::cout << a->nbytes() << " " << a->capacity() << std::endl; // 24, 32

Constructors¶

inline dragon::Tensor::Tensor()¶: Default constructor.

inline explicit dragon::Tensor::Tensor(const string &name)¶: Constructor with the name.

inline explicit dragon::Tensor::Tensor(const vec64_t &dims)¶: Constructor with the int64 dimensions.

inline explicit dragon::Tensor::Tensor(const vec32_t &dims)¶: Constructor with the int32 dimensions.

inline explicit dragon::Tensor::Tensor(const TypeMeta &meta)¶: Constructor with the type meta.

Public Properties¶

axis¶

inline int64_t dragon::Tensor::axis(const int64_t i) const¶: Return a canonical axis

capacity¶

inline size_t dragon::Tensor::capacity() const¶: Return the byte length of memory.

conut¶

inline int64_t dragon::Tensor::count() const¶: Return the number of elements counting along all axes.

conut¶

inline int64_t dragon::Tensor::count(int64_t start) const¶: Return the number of elements counting from given axis.

conut¶

inline int64_t dragon::Tensor::count( int64_t start, int64_t end) const¶: Return the number of elements counting along given axes.

data¶

template<typename T, class Context> inline const T *dragon::Tensor::data()¶: Return the typed data pointer.

dim¶

inline int64_t dragon::Tensor::dim(int64_t i) const¶: Return the dimension of given axis.

dims¶

inline const vec64_t &dragon::Tensor::dims() const¶: Return the dimensions.

empty¶

inline bool dragon::Tensor::empty() const¶: Return whether the number of elements is zero.

memory¶

inline UnifiedMemory *dragon::Tensor::memory( bool required = false, bool owned = false)¶: Return the memory.

memory_state¶

inline UnifiedMemory::State dragon::Tensor::memory_state()¶: Return the memory state.

meta¶

inline const TypeMeta &dragon::Tensor::meta() const¶: Return the type meta.

mutable_data¶

template<typename T, class Context> inline T *dragon::Tensor::mutable_data()¶: Return the typed mutable data pointer.

name¶

inline const string &dragon::Tensor::name() const¶: Return the tensor name.

nbytes¶

inline size_t dragon::Tensor::nbytes() const¶: Return the byte length of all elements.

ndim¶

inline int dragon::Tensor::ndim() const¶: Return the number of dimensions.

raw_data¶

template<class Context> const void *dragon::Tensor::raw_data()¶: Return the raw data pointer.

raw_mutable_data¶

template<class Context> inline void *dragon::Tensor::raw_mutable_data()¶: Return the raw mutable data pointer.

size¶

inline size_t dragon::Tensor::size() const¶: Return the number of elements.

stride¶

inline int64_t dragon::Tensor::stride(int64_t i) const¶: Return the stride of given axis.

strides¶

inline const vec64_t &dragon::Tensor::strides() const¶: Return the strides.

version¶

inline int64_t dragon::Tensor::version() const¶: Return the tensor version.

has_memory¶

inline bool dragon::Tensor::has_memory() const¶: Return whether the memory is set.

has_name¶

inline bool dragon::Tensor::has_name() const¶: Return whether the tensor name is set.

Public Functions¶

CopyFrom¶

template<class Context> inline Tensor *dragon::Tensor::CopyFrom( Tensor &other, Context *ctx)¶: Copy memory from a tensor.

CopyFrom¶

template<typename TensorType, typename VectorType> inline Tensor *dragon::Tensor::CopyFrom(const vector<VectorType> &other)¶: Copy memory from a vector.

CopyTo¶

template<typename TensorType, typename VectorType> inline void dragon::Tensor::CopyTo(vector<VectorType> &dest)¶: Copy memory to a vector.

DimString¶

inline string dragon::Tensor::DimString() const¶: Return a string formatting the tensor dimensions.

DimString¶

static inline string dragon::Tensor::DimString(const vector<int64_t> &dims)¶: Return a string formatting the given dimensions.

IsType¶

template<typename T> inline bool dragon::Tensor::IsType()¶: Return whether the data type is matched.

MapFrom¶

inline Tensor *dragon::Tensor::MapFrom( Tensor *other, size_t offset = 0)¶: Map memory from a tensor.

Reset¶

inline void dragon::Tensor::Reset()¶: Reset tensor to release all resources.

Reshape¶

inline Tensor *dragon::Tensor::Reshape(const vec64_t &dims)¶: Change the dimensions.

ReshapeLike¶

inline Tensor *dragon::Tensor::ReshapeLike(const Tensor &other)¶: Change the dimensions as the other.