Rbm

Collaboration diagram for Rbm:

Namespaces
namespace	cuv::libs::rbm
	Restricted Boltzmann Machine (RBM)

Functions
template<class __value_type , class __memory_space_type , class __memory_layout_type >
void	cuv::libs::rbm::set_binary_sequence (tensor< __value_type, __memory_space_type, __memory_layout_type > &m, const int &start)
	set a matrix to consecutive binary numbers in the columns, starting with the number `start'
template<class __value_type , class __memory_space_type , class __memory_layout_type >
void	cuv::libs::rbm::sigm_temperature (tensor< __value_type, __memory_space_type, __memory_layout_type > &m, const tensor< __value_type, __memory_space_type > &temp)
	apply sigmoid column-wise with the temperature specified for each column
template<class __value_type , class __memory_space_type , class __memory_layout_type >
void	cuv::libs::rbm::set_local_connectivity_in_dense_matrix (tensor< __value_type, __memory_space_type, __memory_layout_type > &m, int patchsize, int vx, int vy, int hx, int hy, int maxdist_from_main_dia=1E6, bool round=false)
	simulate a local connectivity pattern.
template<class __value_type , class __memory_space_type , class __memory_layout_type >
void	cuv::libs::rbm::copy_at_rowidx (tensor< __value_type, __memory_space_type, __memory_layout_type > &dst, const tensor< __value_type, __memory_space_type, __memory_layout_type > &src, const tensor< typename tensor< __value_type, __memory_space_type, __memory_layout_type >::size_type, __memory_space_type, __memory_layout_type > &rowidx, const unsigned int offset)
	copy one matrix into another but only at specified positions.
template<class __value_type , class __memory_space_type , class __memory_layout_type >
void	cuv::libs::rbm::copy_redblack (tensor< __value_type, __memory_space_type, __memory_layout_type > &dst, const tensor< __value_type, __memory_space_type, __memory_layout_type > &src, const unsigned int num_maps, const unsigned int color)
template<class __value_type , class __memory_layout , class __memory_space_type >
void	cuv::libs::rbm::bitflip (tensor< __value_type, __memory_layout, __memory_space_type > &matrix, typename tensor< __value_type, __memory_layout, __memory_space_type >::size_type row)
	Bit-Flip a row of a column-major matrix.

Detailed Description

Function Documentation

template<class __value_type , class __memory_layout , class __memory_space_type >

void cuv::libs::rbm::bitflip	(	tensor< __value_type, __memory_layout, __memory_space_type > &	matrix,
		typename tensor< __value_type, __memory_layout, __memory_space_type >::size_type	row
	)

Bit-Flip a row of a column-major matrix.

Parameters

matrix	Matrix to apply functor on
row	row to flip

changes the matrix such that its m-th row is now (1-original mth row)

template<class __value_type , class __memory_space_type , class __memory_layout_type >

void cuv::libs::rbm::copy_at_rowidx	(	tensor< __value_type, __memory_space_type, __memory_layout_type > &	dst,
		const tensor< __value_type, __memory_space_type, __memory_layout_type > &	src,
		const tensor< typename tensor< __value_type, __memory_space_type, __memory_layout_type >::size_type, __memory_space_type, __memory_layout_type > &	rowidx,
		const unsigned int	offset
	)

copy one matrix into another but only at specified positions.

Parameters

dst	the target matrix (N x M)
src	the source matrix (N x M)
rowidx	contains row indices (M x B). The values at position b*offset+rowidx(m,b) will be copied.
offset	offset (see rowidx)

This kernel can be used to compute serial gibbs updates of a laterally connected RBM hidden layer. Then, offset is the size of one map in the hidden layer, src is the actual result of the calculation and dst is the one where only selected values are changed.

rowidx must be set to indices which are at most as large as the mapsize (<offset, that is). note that for at least some consecutive read operations, rowidx is somewhat "transposed".

template<class __value_type , class __memory_space_type , class __memory_layout_type >

void cuv::libs::rbm::set_binary_sequence	(	tensor< __value_type, __memory_space_type, __memory_layout_type > &	m,
		const int &	start
	)

set a matrix to consecutive binary numbers in the columns, starting with the number `start'

Parameters

m	the target matrix
start	the value of the first column

template<class __value_type , class __memory_space_type , class __memory_layout_type >

void cuv::libs::rbm::set_local_connectivity_in_dense_matrix	(	tensor< __value_type, __memory_space_type, __memory_layout_type > &	m,
		int	patchsize,
		int	vx,
		int	vy,
		int	hx,
		int	hy,
		int	maxdist_from_main_dia = 1E6,
		bool	round = false
	)

simulate a local connectivity pattern.

This simulates the local connectivity in a dense matrix by setting values to zero which are not part of the local connectivity.

assume the lower layer image has dimension vx times vy and the upper layer has size hx times hy. The lower layer has vm maps, the upper layer has hm maps.

Parameters

m	the matrix, should have dimension vx*vy*vm times hx*hy*hm
patchsize	size of local receptive field
vx	as explained above
vy	as explained above
hx
hy
maxdist_from_main_dia	reset everything further than this many maps away from central diagonal
round

template<class __value_type , class __memory_space_type , class __memory_layout_type >

void cuv::libs::rbm::sigm_temperature	(	tensor< __value_type, __memory_space_type, __memory_layout_type > &	m,
		const tensor< __value_type, __memory_space_type > &	temp
	)

apply sigmoid column-wise with the temperature specified for each column

Parameters

m	source and target matrix
temp	the temperature (one value per column)