Tensor Class Reference

#include <tensor.h>

Classes
class	CPUImpl
class	CUDAImpl
class	Impl
class	Shape
class	View

Public Member Functions
	Tensor (const Tensor::Shape &shape, Backend backend=Backend::CPU)
	Constructs a tensor with the given shape, zero-initialized.
	Tensor (const std::initializer_list< size_t > &shape, Backend backend=Backend::CPU)
	Constructs a tensor with the given shape, zero-initialized.
	Tensor (const Tensor::Shape &shape, float value, Backend backend=Backend::CPU)
	Constructs a tensor and fills every element with value.
	Tensor (const std::initializer_list< size_t > &shape, float value, Backend backend=Backend::CPU)
	Constructs a tensor and fills every element with value.
	Tensor (float value, Backend backend=Backend::CPU)
	Constructs a scalar tensor, shape {}.
	Tensor (const Tensor &tensor)
	Copy constructor. Performs a deep copy.
	Tensor (std::unique_ptr< Tensor::Impl > impl, Backend backend=Backend::CPU)
	~Tensor ()
	Destructor.
void	to (Backend newBackend)
	Transfers data to a different backend. No-op if already on that backend.
void	fillAll (float value)
	Fills all elements with value.
void	fillRand ()
	Fills all elements with random values in [-1, 1].
void	print () const
	Prints the tensor to stdout.
void	print (const std::vector< size_t > &position) const
	Prints the block starting at position to stdout.
Tensor::Shape	getShape () const
	Returns the tensor shape.
std::string	getBackendString () const
	Returns "CPU" or "CUDA".
Backend	getBackend () const
	Returns the backend enum.
std::string	getDataString () const
	Returns a string representation of the underlying data.
size_t	getNumElements () const
	Returns the total number of elements.
std::vector< float >	toVector () const
	Copies all elements into a flat vector (row-major order).
void	set (const std::vector< size_t > &position, const Tensor::View &rhs)
	Replaces the block starting at position with the data from rhs.
bool	compare (const Tensor::View &rhs) const
	Returns true if shape and every element matches rhs.
bool	compare (const std::vector< size_t > &position, const Tensor::View &rhs) const
	Returns true if the block at position matches rhs.
Tensor	operator+ (const Tensor::View &rhs) const
	Elementwise addition with a tensor or view.
Tensor	operator- (const Tensor::View &rhs) const
	Elementwise subtraction with a tensor or view.
Tensor	operator* (const Tensor::View &rhs) const
	Elementwise multiplication with a tensor or view.
Tensor	operator/ (const Tensor::View &rhs) const
	Elementwise division by a tensor or view.
Tensor	operator+ (float scalar) const
	Elementwise addition with a pure float.
Tensor	operator- (float scalar) const
	Elementwise subtraction with a pure float.
Tensor	operator* (float scalar) const
	Elementwise multiplication with a pure float.
Tensor	operator/ (float scalar) const
	Elementwise division by a pure float.
void	operator+= (const Tensor::View &rhs)
	In-place elementwise addition with a tensor or view.
void	operator-= (const Tensor::View &rhs)
	In-place elementwise subtraction with a tensor or view.
void	operator*= (const Tensor::View &rhs)
	In-place elementwise multiplication with a tensor or view.
void	operator/= (const Tensor::View &rhs)
	In-place elementwise division by a tensor or view.
Tensor	mean (size_t dim=0) const
	Reduces dimensions [dim, rank) by averaging. Result shape is shape[0:dim].
Tensor	sum (size_t dim=0) const
	Reduces dimensions [dim, rank) by summation. Result shape is shape[0:dim].
Tensor	min (size_t dim=0) const
	Reduces dimensions [dim, rank) by taking the minimum. Result shape is shape[0:dim].
Tensor	max (size_t dim=0) const
	Reduces dimensions [dim, rank) by taking the maximum. Result shape is shape[0:dim].
Tensor	prod (size_t dim=0) const
	Reduces dimensions [dim, rank) by taking the product. Result shape is shape[0:dim].
Tensor	norm () const
	L2 norm (scalar tensor equal to sqrt(sum(x^2))).
Tensor	matmul (const Tensor::View &rhs) const
Tensor::View	operator[] (size_t idx) const
	Indexes into the first dimension, returning a view of the sub-tensor.
Tensor::View	subsample (std::vector< size_t > strides) const
Tensor &	operator= (const Tensor &rhs)
	Copies data from another tensor.
Tensor &	operator= (const Tensor::View &rhs)
	Copies data from a view.
Tensor &	operator= (float scalar)
bool	operator== (const Tensor::View &rhs) const
bool	operator!= (const Tensor::View &rhs) const
Tensor	operator< (const Tensor::View &rhs) const
	Elementwise less than. Returns a tensor of 0.0 / 1.0.
Tensor	operator<= (const Tensor::View &rhs) const
	Elementwise less or equal. Returns a tensor of 0.0 / 1.0.
Tensor	operator> (const Tensor::View &rhs) const
	Elementwise greater than. Returns a tensor of 0.0 / 1.0.
Tensor	operator>= (const Tensor::View &rhs) const
	Elementwise greater or equal. Returns a tensor of 0.0 / 1.0.
Tensor	isClose (const Tensor::View &rhs, float tolerance=1e-5f) const

Friends
Tensor	operator+ (float scalar, const Tensor &rhs)
	Elementwise addition of a pure float and a tensor.
Tensor	operator- (float scalar, const Tensor &rhs)
	Elementwise subtraction of a tensor from a pure float.
Tensor	operator* (float scalar, const Tensor &rhs)
	Elementwise multiplication of a pure float and a tensor.
Tensor	operator/ (float scalar, const Tensor &rhs)
	Elementwise division of a pure float by a tensor.

Detailed Description

Multi-dimensional array with backend specific implementation.

A Tensor owns a contiguous block of float data managed by a backend specific Impl object. Elements can be accessed through Tensor::View, which describes a sub region via offset, shape, and stride layout.

Constructor & Destructor Documentation

Tensor()

Tensor::Tensor	(	std::unique_ptr< Tensor::Impl >	impl,
		Backend	backend = Backend::CPU
	)

Constructs a tensor by taking ownership of a backend implementation.

Parameters

impl	Backend implementation, ownership is transferred.

Member Function Documentation

isClose()

Tensor Tensor::isClose	(	const Tensor::View &	rhs,
		float	tolerance = 1e-5f
	)		const

Elementwise closeness check within tolerance. Returns a tensor of 0.0 / 1.0.

Parameters

tolerance

Maximum absolute or relative difference (default: 1e-5).

matmul()

Tensor Tensor::matmul

(

const Tensor::View &

rhs

)

const

Matrix multiplication. Inputs must be 2D or 3D tensors. 2D: (N, M) @ (M, K) => (N, K).

3D: (B, N, M) @ (B, M, K) => (B, N, K).

Batch dims must be broadcastable, match or be 1.

operator!=()

bool Tensor::operator!=

(

const Tensor::View &

rhs

)

const

Returns true if any element differs from rhs.

Note

Exact match, which is unstable for floats. Consider using .isClose()

operator=()

Tensor & Tensor::operator=

(

float

scalar

)

Assigns a pure float to a tensor.

Note

Only works on scalar-shaped tensors, e.g tensors with only 1 element.

operator==()

bool Tensor::operator==

(

const Tensor::View &

rhs

)

const

Returns true if all elements equal those in rhs.

Note

Exact match, which is unstable for floats. Consider using .isClose()

subsample()

Tensor::View Tensor::subsample

(

std::vector< size_t >

strides

)

const

Strided sub-sampling view. Views every strides[i]-th element along dim i.

Parameters

strides

Length must match rank of tensor or be scalar 0.

---

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

• include/nforge/core/tensor.h
• src/core/tensor.cpp