Tensor Class Reference

Class to represent multidimensional arrays. More...

#include <tensor.h>

Public Member Functions
	Tensor ()
	Initialize an empty tensor.
	Tensor (const Tensor &other)
	The copy constructor. More...
	Tensor (std::vector< size_t > shape)
	Tensor constructor with shape given directly by a vector. More...
	Tensor (std::vector< size_t > shape, DataLocation location)
template<typename... Args>
	Tensor (Args... args)
	Construct a tensor based on the passed shape. More...
Tensor &	operator= (const Tensor &other)
	The assignment operator. More...
std::shared_ptr< Tensor >	copy () const
void	move (DataLocation target)
	Move the tensor to the designated destination. More...
void	useGrad ()
void	backward ()
template<typename... Args>
float &	operator() (Args... args)
	Method to access an element of the tensor at a specific method. More...
	~Tensor ()
	The destructor.

Static Public Member Functions
static Tensor	construct1d (const std::vector< float > &data)
	Static method to easily initialize a 1D tensor with given data. More...
static Tensor	construct2d (const std::vector< std::vector< float >> &data)
	Static method to easily initialize a 2D tensor with given data. More...

Public Attributes
std::vector< size_t >	shape
	Store the shape of the tensor.
size_t	size
	Store the total size of the tensor. More...
DataLocation	location
	The location of the tensor. More...
float *	data
	The data stored by the tensor.
Session	session
	Session object containing information about current session.
bool	requiresGrad
std::shared_ptr< BackwardFunction >	gradFunction
std::shared_ptr< Tensor >	grad

Private Member Functions
void	computeSize ()
	Compute the total size of the tensor based on its shape.
size_t	findEffectiveAddress (const std::vector< size_t > &index, size_t depth) const
	Finds the address of an element in the flattened data array given its index in non-flattened tensor. More...
void	verifyIndex (const std::vector< size_t > &index) const
	Verify that an index of an element is within the shape of the tensor. More...

Detailed Description

Class to represent multidimensional arrays.

Constructor & Destructor Documentation

Tensor() [1/3]

Tensor::Tensor

(

const Tensor &

other

)

The copy constructor.

Parameters

other

The tensor based on which this one should be initialized.

Tensor() [2/3]

Tensor::Tensor ( std::vector< size_t >  shape )

explicit

Tensor constructor with shape given directly by a vector.

Parameters

shape

The shape of the vector to create.

Tensor() [3/3]

template<typename... Args>

Tensor::Tensor ( Args...  args )

inlineexplicit

Construct a tensor based on the passed shape.

For example Tensor(2, 2, 3) will initialize a 3D tensor with two 2x3 matrices.

Parameters

args	The shapes of consecutive dimensions.

Member Function Documentation

construct1d()

Tensor Tensor::construct1d ( const std::vector< float > &  data )

static

Static method to easily initialize a 1D tensor with given data.

Parameters

data	The data based on which a tensor should be constructed.

Returns

The constructed tensor.

construct2d()

Tensor Tensor::construct2d ( const std::vector< std::vector< float >> &  data )

static

Static method to easily initialize a 2D tensor with given data.

Parameters

data	The data based on which a tensor should be constructed.

Returns

The constructed tensor.

findEffectiveAddress()

size_t Tensor::findEffectiveAddress ( const std::vector< size_t > &  index, size_t  depth  ) const

private

Finds the address of an element in the flattened data array given its index in non-flattened tensor.

Since the method works recursively, it also takes the depth parameter, which provides information about which dimension is currently taken into account when computing the index.

Parameters

index	The multidimensional index of the element to access.
depth	The dimension that is currently considered in the recursive call.

Returns

The address of the element in the flattened data array.

move()

void Tensor::move

(

DataLocation

target

)

Move the tensor to the designated destination.

This involves copying the data to the new location and releasing memory from the old location.

Parameters

target

The destination to move the tensor to.

operator()()

template<typename... Args>

float& Tensor::operator() ( Args...  args )

inline

Method to access an element of the tensor at a specific method.

This method should not be used in performance critical operations. In these cases a direct access to Tensor::data will be more appropriate (as it becomes easier for the compiler to optimize such code).

Parameters

args	The index of the element to access.

Returns

A reference to the requested element.

operator=()

Tensor & Tensor::operator=

(

const Tensor &

other

)

The assignment operator.

This releases all the current memory and copies all the information of the other tensor.

Parameters

other

The other tensor to assign this one to.

Returns

Assigned tensor. Always returns *this.

verifyIndex()

void Tensor::verifyIndex ( const std::vector< size_t > &  index ) const

private

Verify that an index of an element is within the shape of the tensor.

Parameters

index

The index of the element that is being accessed.

Member Data Documentation

location

DataLocation Tensor::location

The location of the tensor.

Can be either HOST or DEVICE. See DataLocation for more details.

size

size_t Tensor::size

Store the total size of the tensor.

The size is equal to the number of elements that can be contained in the tensor. It is computed in the Tensor::computeSize() method.

---

The documentation for this class was generated from the following files:

• include/tensor.h
• src/math/tensor.cpp