BLAS1 -- Vector/Vector operations

Collaboration diagram for BLAS1 -- Vector/Vector operations:

Modules
	Pointwise functors on vectors
	Functors reducing a vector to a scalar

Functions
template<class __value_type , class __memory_space_type , class S >
void	cuv::rprop (tensor< __value_type, __memory_space_type > &W, tensor< __value_type, __memory_space_type > &dW, tensor< S, __memory_space_type > &dW_old, tensor< __value_type, __memory_space_type > &rate, const float &decay=0.0f, const float &sparsedecay=0.0f)
	Does a gradient descent step using the "RPROP" algorithm.
template<class __value_type , class __memory_space_type , class S >
void	cuv::rprop (tensor< __value_type, __memory_space_type, column_major > &W, tensor< __value_type, __memory_space_type, column_major > &dW, tensor< S, __memory_space_type, column_major > &dW_old, tensor< __value_type, __memory_space_type, column_major > &rate, const float &decay=0.0f, const float &sparsedecay=0.0f)
	This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts. casting column major to row major since working on linear memory anyway.
template<class __value_type , class __memory_space_type >
void	cuv::learn_step_weight_decay (tensor< __value_type, __memory_space_type > &W, const tensor< __value_type, __memory_space_type > &dW, const float &learnrate, const float &decay=0.0f, const float &sparsedecay=0.0f)
	Do a step of gradient descent with optional weight decay.
template<class V , class M >
void	cuv::learn_step_weight_decay_momentum (tensor< V, M > &W, tensor< V, M > &momentum, const tensor< V, M > &dW, const float &learnrate, const float &momentum_weight=0.9, const float &decay=0.0f, const float &sparsedecay=0.0f)
	Same as learn_step_weight_decay, but with momentum.
template<class __value_type , class __memory_space_type >
void	cuv::learn_step_weight_decay (tensor< __value_type, __memory_space_type, column_major > &W, const tensor< __value_type, __memory_space_type, column_major > &dW, const float &learnrate, const float &decay=0.0f, const float &sparsedecay=0.0f)
	This is an overloaded member function, provided for convenience. It differs from the above function only in what argument(s) it accepts. casting column major to row major since working on linear memory anyway.

Detailed Description

Function Documentation

template<class __value_type , class __memory_space_type >

void cuv::learn_step_weight_decay	(	tensor< __value_type, __memory_space_type > &	W,
		const tensor< __value_type, __memory_space_type > &	dW,
		const float &	learnrate,
		const float &	decay = 0.0f,
		const float &	sparsedecay = 0.0f
	)

Do a step of gradient descent with optional weight decay.

Parameters

W	Destination matrix
dW	Direction of gradient descent. Vector of same size as W.
learnrate	Scalar learnreate
decay	Scalar L2 weight decay (cost) parameter
sparsedecay	Scalar L1 weight decay (cost) parameter

Calculates W = (1-decay*learnrate) * W + learnrate * dW

template<class V , class M >

void cuv::learn_step_weight_decay_momentum	(	tensor< V, M > &	W,
		tensor< V, M > &	momentum,
		const tensor< V, M > &	dW,
		const float &	learnrate,
		const float &	momentum_weight = 0.9,
		const float &	decay = 0.0f,
		const float &	sparsedecay = 0.0f
	)

Same as learn_step_weight_decay, but with momentum.

Parameters

W	Destination matrix
momentum	The accumulated momentum (IN and OUT)
dW	Direction of gradient descent. Vector of same size as W.
learnrate	Scalar learnreate
momentum_weight	how strong to rely on accumulated momentum
decay	Scalar L2 weight decay (cost) parameter
sparsedecay	Scalar L1 weight decay (cost) parameter

template<class __value_type , class __memory_space_type , class S >

void cuv::rprop	(	tensor< __value_type, __memory_space_type > &	W,
		tensor< __value_type, __memory_space_type > &	dW,
		tensor< S, __memory_space_type > &	dW_old,
		tensor< __value_type, __memory_space_type > &	rate,
		const float &	decay = 0.0f,
		const float &	sparsedecay = 0.0f
	)

Does a gradient descent step using the "RPROP" algorithm.

Parameters

W	Destination tensor
dW	Direction of gradient descent. Vector of same size as W.
dW_old	Direction of gradient descent in privious step. Vector of same size as W.
rate	Vector of same size as W containing separate learnrates for each entry.
decay	Scalar L2 weight decay (cost) parameter
sparsedecay	Scalar L1 weight decay (cost) parameter Updates W according to the "RPROP" algorithm. Calculates W = (1-decay*rate)*W + rate * W where all multiplications are pointwise. Also rate and dW_old are updated at each step.