#include <cutout.hpp>

Inheritance diagram for lbann::transform::cutout:

Collaboration diagram for lbann::transform::cutout:

Public Member Functions
	cutout (size_t num_holes, size_t length)

transform *	copy () const override

std::string	get_type () const override

void	apply (utils::type_erased_matrix &data, std::vector< size_t > &dims) override

Public Member Functions inherited from lbann::transform::transform
	transform ()=default

	transform (const transform &)=default

transform &	operator= (const transform &)=default

virtual	~transform ()=default

virtual description	get_description () const

virtual bool	supports_non_inplace () const

virtual void	apply (utils::type_erased_matrix &data, CPUMat &out, std::vector< size_t > &dims)

Private Attributes
size_t	m_num_holes

size_t	m_length

Additional Inherited Members
Static Protected Member Functions inherited from lbann::transform::transform
static float	get_uniform_random (float a, float b)

static bool	get_bool_random (float p)

static El::Int	get_uniform_random_int (El::Int a, El::Int b)

Detailed Description

Cutout data augmentation which randomly masks out square regions of input.

See:

DeVries and Taylor. "Improved Regularization of Convolutional Neural
Networks with Cutout". arXiv preprint arXiv:1708.04552 (2017).

This will randomly select a center pixel for each square and set all pixels within that square to 0. It is permissible for portions of the masks to lie outside of the image.

Normalization about 0 should be applied after applying cutout.

Definition at line 51 of file transforms/vision/cutout.hpp.

Constructor & Destructor Documentation

◆ cutout()

lbann::transform::cutout::cutout	(	size_t	num_holes,
		size_t	length
	)

inline

Cutout with a given number of squares of a given size.

Parameters

num_holes	Number of squares to mask out (must be positive).
length	Length of a side of the square (must be positive).

Definition at line 59 of file transforms/vision/cutout.hpp.

Here is the caller graph for this function:

Member Function Documentation

◆ apply()

void lbann::transform::cutout::apply	(	utils::type_erased_matrix &	data,
		std::vector< size_t > &	dims
	)

overridevirtual

Apply the transform to data.

Parameters

data	The input data to transform, which is modified in-place. The matrix shuold be contiguous.
dims	The dimensions of the data tensor. For "plain data", dims should have one entry, giving its size. For images, dims should have three entries: channels, height, width.