mlpack::nn::SparseAutoencoder Class Reference

A sparse autoencoder is a neural network whose aim to learn compressed representations of the data, typically for dimensionality reduction, with a constraint on the activity of the neurons in the network. More...

#include <sparse_autoencoder.hpp>

Public Member Functions
template<typename OptimizerType = ens::L_BFGS>
	SparseAutoencoder (const arma::mat &data, const size_t visibleSize, const size_t hiddenSize, const double lambda=0.0001, const double beta=3, const double rho=0.01, OptimizerType optimizer=OptimizerType())
	Construct the sparse autoencoder model with the given training data. More...
template<typename OptimizerType , typename... CallbackTypes>
	SparseAutoencoder (const arma::mat &data, const size_t visibleSize, const size_t hiddenSize, const double lambda, const double beta, const double rho, OptimizerType optimizer, CallbackTypes &&... callbacks)
	Construct the sparse autoencoder model with the given training data. More...
void	GetNewFeatures (arma::mat &data, arma::mat &features)
	Transforms the provided data into the representation learned by the sparse autoencoder. More...
void	Sigmoid (const arma::mat &x, arma::mat &output) const
	Returns the elementwise sigmoid of the passed matrix, where the sigmoid function of a real number 'x' is [1 / (1 + exp(-x))]. More...
void	VisibleSize (const size_t visible)
	Sets size of the visible layer.
size_t	VisibleSize () const
	Gets size of the visible layer.
void	HiddenSize (const size_t hidden)
	Sets size of the hidden layer.
size_t	HiddenSize () const
	Gets the size of the hidden layer.
void	Lambda (const double l)
	Sets the L2-regularization parameter.
double	Lambda () const
	Gets the L2-regularization parameter.
void	Beta (const double b)
	Sets the KL divergence parameter.
double	Beta () const
	Gets the KL divergence parameter.
void	Rho (const double r)
	Sets the sparsity parameter.
double	Rho () const
	Gets the sparsity parameter.

Detailed Description

A sparse autoencoder is a neural network whose aim to learn compressed representations of the data, typically for dimensionality reduction, with a constraint on the activity of the neurons in the network.

Sparse autoencoders can be stacked together to learn a hierarchy of features, which provide a better representation of the data for classification. This is a method used in the recently developed field of deep learning. More technical details about the model can be found on the following webpage:

http://deeplearning.stanford.edu/wiki/index.php/UFLDL_Tutorial

An example of how to use the interface is shown below:

arma::mat data; // Data matrix.
const size_t vSize = 64; // Size of visible layer, depends on the data.
const size_t hSize = 25; // Size of hidden layer, depends on requirements.

// Train the model using default options.
SparseAutoencoder encoder1(data, vSize, hSize);

const size_t numBasis = 5; // Parameter required for L-BFGS algorithm.
const size_t numIterations = 100; // Maximum number of iterations.

// Use an instantiated optimizer for the training.
SparseAutoencoderFunction saf(data, vSize, hSize);
L_BFGS<SparseAutoencoderFunction> optimizer(saf, numBasis, numIterations);
SparseAutoencoder<L_BFGS> encoder2(optimizer);

arma::mat features1, features2; // Matrices for storing new representations.

// Get new representations from the trained models.
encoder1.GetNewFeatures(data, features1);
encoder2.GetNewFeatures(data, features2);

This implementation allows the use of arbitrary mlpack optimizers via the OptimizerType template parameter.

Constructor & Destructor Documentation

SparseAutoencoder() [1/2]

template<typename OptimizerType >

mlpack::nn::SparseAutoencoder::SparseAutoencoder	(	const arma::mat &	data,
		const size_t	visibleSize,
		const size_t	hiddenSize,
		const double	lambda = 0.0001,
		const double	beta = 3,
		const double	rho = 0.01,
		OptimizerType	optimizer = OptimizerType()
	)

Construct the sparse autoencoder model with the given training data.

This will train the model. The parameters 'lambda', 'beta' and 'rho' can be set optionally. Changing these parameters will have an effect on regularization and sparsity of the model.

Template Parameters

OptimizerType

The optimizer to use.

Parameters

data	Input data with each column as one example.
visibleSize	Size of input vector expected at the visible layer.
hiddenSize	Size of input vector expected at the hidden layer.
lambda	L2-regularization parameter.
beta	KL divergence parameter.
rho	Sparsity parameter.
optimizer	Desired optimizer.

SparseAutoencoder() [2/2]

template<typename OptimizerType , typename... CallbackTypes>

mlpack::nn::SparseAutoencoder::SparseAutoencoder	(	const arma::mat &	data,
		const size_t	visibleSize,
		const size_t	hiddenSize,
		const double	lambda,
		const double	beta,
		const double	rho,
		OptimizerType	optimizer,
		CallbackTypes &&...	callbacks
	)

Construct the sparse autoencoder model with the given training data.

This will train the model. The parameters 'lambda', 'beta' and 'rho' can be set optionally. Changing these parameters will have an effect on regularization and sparsity of the model.

Template Parameters

OptimizerType	The optimizer to use.
CallbackTypes	Types of Callback Functions.

Parameters

data	Input data with each column as one example.
visibleSize	Size of input vector expected at the visible layer.
hiddenSize	Size of input vector expected at the hidden layer.
lambda	L2-regularization parameter.
beta	KL divergence parameter.
rho	Sparsity parameter.
optimizer	Desired optimizer.
callbacks	Callback function for ensmallen optimizer OptimizerType. See https://www.ensmallen.org/docs.html#callback-documentation.

Member Function Documentation

GetNewFeatures()

void mlpack::nn::SparseAutoencoder::GetNewFeatures	(	arma::mat &	data,
		arma::mat &	features
	)

Transforms the provided data into the representation learned by the sparse autoencoder.

The function basically performs a feedforward computation using the learned weights, and returns the hidden layer activations.

Parameters

data	Matrix of the provided data.
features	The hidden layer representation of the provided data.

Sigmoid()

void mlpack::nn::SparseAutoencoder::Sigmoid ( const arma::mat &  x, arma::mat &  output  ) const

inline

Returns the elementwise sigmoid of the passed matrix, where the sigmoid function of a real number 'x' is [1 / (1 + exp(-x))].

Parameters

x	Matrix of real values for which we require the sigmoid activation.
output	Output matrix.

---

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

• src/mlpack/methods/sparse_autoencoder/sparse_autoencoder.hpp
• src/mlpack/methods/sparse_autoencoder/sparse_autoencoder.cpp
• src/mlpack/methods/sparse_autoencoder/sparse_autoencoder_impl.hpp