aureservoir Namespace Reference

Data Structures
class	AUExcept
class	CalcDelay
	template class for delay calculation More...
class	DelayLine
	template class for a signal delay line More...
class	ESN
	class for a basic Echo State Network More...
class	BPFilter
	simple bandpass filter based on an exponential moving average More...
class	IIRFilter
	general IIR filter implemented in transposed direct form 2 More...
class	SerialIIRFilter
	serie of IIR Filters More...
class	InitBase
	abstract base class for initialization algorithms More...
class	InitStd
	standard initialization as described in Jaeger's initial paper More...
class	SimBase
	abstract base class for simulation algorithms More...
class	SimStd
	standard simulation algorithm as in Jaeger's initial paper More...
class	SimLI
	algorithm with leaky integrator neurons More...
class	SimBP
	algorithm with bandpass style neurons as in Wustlich and Siewert More...
class	SimFilter
	algorithm with general IIR-Filter neurons More...
class	SimFilter2
	algorithm with IIR-Filter before neuron nonlinearity More...
class	SimFilterDS
	IIR-Filter neurons with additional delay&sum readout. More...
class	SimSquare
	algorithm with additional squared state updates More...
class	TrainBase
	abstract base class for training algorithms More...
class	TrainPI
	offline trainig algorithm using the pseudo inverse More...
class	TrainLS
	offline trainig algorithm using the least square solution More...
class	TrainRidgeReg
	offline algorithm with Ridge Regression / Tikhonov Regularization More...
class	TrainDSPI
	offline algorithm for delay&sum readout with PI More...
struct	SPMatrix
	typedef trait class of a real sparse matrix with type T More...
struct	DEMatrix
	typedef trait class of a real dense matrix with type T More...
struct	DEVector
	typedef trait class of a real dense vector with type T More...
struct	CDEVector
	typedef trait class of a complex dense vector with type T More...
class	Rand
	template class for random number generation More...
tanh2 activation functions
DEVector< double >::Type	tanh2_a_
DEVector< double >::Type	tanh2_b_
template<typename T>
void	act_tanh2 (T *data, int size)
template<typename T>
void	act_invtanh2 (T *data, int size)
Enumerations
enum	ActivationFunction { ACT_LINEAR, ACT_TANH, ACT_TANH2, ACT_SIGMOID }
enum	InitAlgorithm { INIT_STD }
enum	InitParameter {   CONNECTIVITY, ALPHA, IN_CONNECTIVITY, IN_SCALE,   IN_SHIFT, FB_CONNECTIVITY, FB_SCALE, FB_SHIFT,   LEAKING_RATE, TIKHONOV_FACTOR, DS_USE_CROSSCORR, DS_USE_GCC,   DS_MAXDELAY, IP_LEARNRATE, IP_MEAN, IP_VAR }
enum	SimAlgorithm {   SIM_STD, SIM_SQUARE, SIM_LI, SIM_BP,   SIM_FILTER, SIM_FILTER2, SIM_FILTER_DS }
enum	TrainAlgorithm { TRAIN_PI, TRAIN_LS, TRAIN_RIDGEREG, TRAIN_DS_PI }
Functions
void	set_denormal_flags () throw (AUExcept)
void	denormals_add_dc (float *data, int size)
void	denormals_add_dc (double *data, int size)
double	stringToDouble (const string &s) throw (AUExcept)
int	stringToInt (const string &s) throw (AUExcept)
linear activation functions
template<typename T>
void	act_linear (T *data, int size)
template<typename T>
void	act_invlinear (T *data, int size)
tanh activation functions
template<typename T>
void	act_tanh (T *data, int size)
template<typename T>
void	act_invtanh (T *data, int size)
sigmoid activation functions
template<typename T>
void	act_sigmoid (T *data, int size)
template<typename T>
void	act_invsigmoid (T *data, int size)
FFT routines using FFTW
void	rfft (const DEVector< double >::Type &x, CDEVector< double >::Type &X, int fftsize)
void	rfft (const DEVector< float >::Type &x, CDEVector< float >::Type &X, int fftsize)
void	irfft (CDEVector< double >::Type &X, DEVector< double >::Type &x)
void	irfft (CDEVector< float >::Type &X, DEVector< float >::Type &x)
Variables
const float	SINGLE_DENORMAL_DC = 1.0E-25
const double	DOUBLE_DENORMAL_DC = 1.0E-30

---

Enumeration Type Documentation

enum aureservoir::ActivationFunction

all possible activation functions for reservoir and output neurons

Enumerator:

ACT_LINEAR	linear activation function
ACT_TANH	tanh activation function
ACT_TANH2	tanh activation function with local slope and bias
ACT_SIGMOID	sigmoid activation function

enum aureservoir::InitAlgorithm

all possible initialization algorithms

Enumerator:

INIT_STD

standard initialization, See also: class InitStd

enum aureservoir::InitParameter

possible parameters of the initialization algorithms

Note:

not every algorithm must use all of them !

Enumerator:

CONNECTIVITY	connectivity of the reservoir weight matrix
ALPHA	spectral radius of the reservoir weight matrix
IN_CONNECTIVITY	connectivity of the input weight matrix
IN_SCALE	scale input weight matrix random vaules
IN_SHIFT	shift input weight matrix random vaules
FB_CONNECTIVITY	connectivity of the feedback weight matrix
FB_SCALE	scale feedback weight matrix random vaules
FB_SHIFT	shift feedback weight matrix random vaules
LEAKING_RATE	leaking rate for Leaky Integrator ESNs
TIKHONOV_FACTOR	regularization factor for TrainRidgeReg
DS_USE_CROSSCORR	use simple cross-correlation for delay calculation
DS_USE_GCC	use generalized cross-correlation for delay calculation
DS_MAXDELAY	maximum delay for delay&sum readout
IP_LEARNRATE	learnrate for Gaussian-IP reservoir adaptation
IP_MEAN	desired mean for Gaussian-IP reservoir adaptation
IP_VAR	desired variance for Gaussian-IP reservoir adaptation

enum aureservoir::SimAlgorithm

all possible simulation algorithms

See also:

class SimStd

Enumerator:

SIM_STD	standard simulation See also: class SimStd
SIM_SQUARE	additional squared state updates See also: class SimSquare
SIM_LI	simulation with leaky integrator neurons See also: class SimLI
SIM_BP	simulation with bandpass neurons See also: class SimBP
SIM_FILTER	simulation with IIR-Filter neurons See also: class SimFilter
SIM_FILTER2	IIR-Filter before nonlinearity. See also: class SimFilter2
SIM_FILTER_DS	with Delay&Sum Readout See also: class SimFilterDS

enum aureservoir::TrainAlgorithm

all possible training algorithms

Enumerator:

TRAIN_PI	offline, pseudo inverse based See also: class TrainPI
TRAIN_LS	offline least square algorithm, See also: class TrainLS
TRAIN_RIDGEREG	with ridge regression, See also: class TrainRidgeReg
TRAIN_DS_PI	trains a delay&sum readout with PI See also: class TrainDSPI

---

Function Documentation

template<typename T>

void aureservoir::act_invlinear	(	T *	data,
		int	size
	)			[inline]

inverse linear activation function, performed on each element

Parameters:

	data	pointer to the data
	size	of the data

template<typename T>

void aureservoir::act_invsigmoid	(	T *	data,
		int	size
	)			[inline]

inverse sigmoid activation function, performed on each element: y(x) = ln( 1/x - 1 )

template<typename T>

void aureservoir::act_invtanh	(	T *	data,
		int	size
	)			[inline]

inverse tanh activation function, performed on each element

Parameters:

	data	pointer to the data
	size	of the data

template<typename T>

void aureservoir::act_invtanh2	(	T *	data,
		int	size
	)			[inline]

inverse tanh2 activation function this means the following: y(x) = (atanh(x) - b) / a

Parameters:

	data	pointer to the data
	size	of the data

template<typename T>

void aureservoir::act_linear	(	T *	data,
		int	size
	)			[inline]

linear activation function, performed on each element

Parameters:

	data	pointer to the data
	size	of the data

template<typename T>

void aureservoir::act_sigmoid	(	T *	data,
		int	size
	)			[inline]

sigmoid activation function, performed on each element: y(x) = 1 / (1 + exp(x))

Parameters:

	data	pointer to the data
	size	of the data

template<typename T>

void aureservoir::act_tanh	(	T *	data,
		int	size
	)			[inline]

tanh activation function, performed on each element

Parameters:

	data	pointer to the data
	size	of the data

template<typename T>

void aureservoir::act_tanh2	(	T *	data,
		int	size
	)			[inline]

tanh2 activation function with local slope a and bias b this means the following: y(x) = tanh( a*x + b ) where a and b are vetors with same size as the data

Parameters:

	data	pointer to the data
	size	of the data

void aureservoir::denormals_add_dc	(	double *	data,
		int	size
	)			[inline]

adds a constant value to prevent going to denormal mode

Parameters:

	data	pointer to double precision data
	size	of the data

See also:

http://www.musicdsp.org/files/other001.txt

void aureservoir::denormals_add_dc	(	float *	data,
		int	size
	)			[inline]

adds a constant value to prevent going to denormal mode

Parameters:

	data	pointer to single precision data
	size	of the data

See also:

http://www.musicdsp.org/files/other001.txt

void aureservoir::irfft	(	CDEVector< float >::Type &	X,
		DEVector< float >::Type &	x
	)

calculates inverse real fft in single precision

Parameters:

	X	complex frequency domain input vector
	x	real IFFT output vector, will be resized to correct size

void aureservoir::irfft	(	CDEVector< double >::Type &	X,
		DEVector< double >::Type &	x
	)

calculates inverse real fft in double precision

Parameters:

	X	complex frequency domain input vector
	x	real IFFT output vector, will be resized to correct size

void aureservoir::rfft	(	const DEVector< float >::Type &	x,
		CDEVector< float >::Type &	X,
		int	fftsize
	)

calculates real fft with zero padding in single precision

Parameters:

	x	real input vector
	X	complex FFT output vector, will be resized to fftsize/2+1
	fftsize	fftsize, x will be zero-padded to this size

void aureservoir::rfft	(	const DEVector< double >::Type &	x,
		CDEVector< double >::Type &	X,
		int	fftsize
	)

calculates real fft with zero padding in double precision

Parameters:

	x	real input vector
	X	complex FFT output vector, will be resized to fftsize/2+1
	fftsize	fftsize, x will be zero-padded to this size

void aureservoir::set_denormal_flags

(

)

throw (AUExcept) [inline]

deactivate pentium4 denormals

This deactivates denormals on pentium4 processors, which speeds up quite a lot. However, this only works for the SSE unit and is not supported by older models, so all used libraries have to be compiled with -mfpmath=sse -msse flags. To turn denormals off on SSE, we turn on the Denormals Are Zero and Flush to Zero (DAZ and FZ) bits in the MXCSR register.

See also:

http://developer.apple.com/documentation/Performance/Conceptual/Accelerate_sse_migration/migration_sse_translation/chapter_4_section_2.html

Note:

For some strange reasons this function must be inline, otherwise I get linker problems !

double aureservoir::stringToDouble

(

const string &

s

)

throw (AUExcept) [inline]

specialization for complex

converts a value from a string to a double used to set parameters from strings

int aureservoir::stringToInt

(

const string &

s

)

throw (AUExcept) [inline]

converts a value from a string to a integer used to set parameters from strings

---

Variable Documentation

const double aureservoir::DOUBLE_DENORMAL_DC = 1.0E-30

DC offset to prevent denormals for double precision

const float aureservoir::SINGLE_DENORMAL_DC = 1.0E-25

DC offset to prevent denormals for single precision

DEVector<double>::Type aureservoir::tanh2_a_

slope vector for tanh2

DEVector<double>::Type aureservoir::tanh2_b_

bias vector for tanh2

---