DNF Class Reference

Deep Neuronal Filter https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0277974. More...

#include <dnf_torch.h>

Public Types
enum	ActMethod { Act_Sigmoid = 1 , Act_Tanh = 2 , Act_ReLU = 3 , Act_NONE = 0 }
	Options for activation functions of all neurons in the network.

Public Member Functions
	DNF (const int nLayers, const int nTaps, const ActMethod am=Act_Tanh, const bool tryGPU=false)
	Constructor which sets up the delay lines, network layers and also calculates the number of neurons per layer so that the final layer always just has one neuron.
void	setLearningRate (float mu)
	Sets the learning rate of the entire network.
float	filter (const float signal, const float noise)
	Realtime sample by sample filtering operation.
int	getSignalDelaySteps () const
	Returns the length of the delay line which delays the signal polluted with noise.
float	getDelayedSignal () const
	Returns the delayed with noise polluted signal by the delay indicated by getSignalDelaySteps().
float	getRemover () const
	Returns the remover signal.
float	getOutput () const
	Returns the output of the DNF: the noise free signal.
const std::vector< float >	getLayerWeightDistances () const
	Gets the weight distances per layer.
float	getWeightDistance () const
	Gets the overall weight distsance.
const torch::Device	getTorchDevice () const
	Gets the torch device for example to determine if the GPU is being used.
const Net	getModel () const
	Gets the torch model, for example, to read out the weights.

Static Public Attributes
static constexpr double	xavierGain = 0.01
	Xavier gain for the weight init.

Detailed Description

Deep Neuronal Filter https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0277974.

Constructor & Destructor Documentation

DNF()

DNF::DNF	(	const int	nLayers,
		const int	nTaps,
		const ActMethod	am = Act_Tanh,
		const bool	tryGPU = false
	)

Constructor which sets up the delay lines, network layers and also calculates the number of neurons per layer so that the final layer always just has one neuron.

Parameters

nLayers	Number of layers
nTaps	Number of taps for the delay line feeding into the 1st layer
am	The activation function for the neurons. Default is tanh.
tryGPU	Does the learning on the GPU if available.

Member Function Documentation

filter()

float DNF::filter	(	const float	signal,
		const float	noise
	)

Realtime sample by sample filtering operation.

Parameters

signal	The signal contaminated with noise. Should be less than one.
noise	The reference noise. Should be less than one.

Returns

The filtered signal where the noise has been removed by the DNF.

getDelayedSignal()

float DNF::getDelayedSignal ( ) const

inline

Returns the delayed with noise polluted signal by the delay indicated by getSignalDelaySteps().

Returns

The delayed noise polluted signal sample.

getLayerWeightDistances()

const std::vector< float > DNF::getLayerWeightDistances

(

)

const

Gets the weight distances per layer.

Returns

The Eucledian weight distance in relation to the initial weights.

getOutput()

float DNF::getOutput ( ) const

inline

Returns the output of the DNF: the noise free signal.

Returns

The current output of the DNF which is idential to filter().

getRemover()

float DNF::getRemover ( ) const

inline

Returns the remover signal.

Returns

The current remover signal sample.

getSignalDelaySteps()

int DNF::getSignalDelaySteps ( ) const

inline

Returns the length of the delay line which delays the signal polluted with noise.

Returns

Number of delay steps in samples.

getWeightDistance()

float DNF::getWeightDistance

(

)

const

Gets the overall weight distsance.

Returns

The sum of all layer weight distances.

setLearningRate()

void DNF::setLearningRate

(

float

mu

)

Sets the learning rate of the entire network.

It can be set any time during learning. Setting it to zero disables learning / adaptation.

Parameters

mu	Learning rate

---

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

• dnf_torch.h