mic::neural_nets::optimization::AdaDelta< eT > Class Template Reference

Update using AdaDelta - adaptive gradient descent with running average E[g^2] and E[d^2]. More...

#include <AdaDelta.hpp>

Inheritance diagram for mic::neural_nets::optimization::AdaDelta< eT >:

Collaboration diagram for mic::neural_nets::optimization::AdaDelta< eT >:

Public Member Functions
	AdaDelta (size_t rows_, size_t cols_, eT decay_=0.9, eT eps_=1e-8)
mic::types::MatrixPtr< eT >	calculateUpdate (mic::types::MatrixPtr< eT > x_, mic::types::MatrixPtr< eT > dx_, eT learning_rate_)
Public Member Functions inherited from mic::neural_nets::optimization::OptimizationFunction< eT >
	OptimizationFunction ()
virtual	~OptimizationFunction ()
	Virtual destructor - empty. More...
virtual void	update (mic::types::MatrixPtr< eT > p_, mic::types::MatrixPtr< eT > dp_, eT learning_rate_, eT decay_=0.0)
virtual void	update (mic::types::MatrixPtr< eT > p_, mic::types::MatrixPtr< eT > x_, mic::types::MatrixPtr< eT > y_, eT learning_rate_=0.001)

Protected Attributes
eT	decay
	Decay ratio, similar to momentum. More...
eT	eps
	Smoothing term that avoids division by zero. More...
mic::types::MatrixPtr< eT >	EG
	Decaying average of the squares of gradients up to time t ("diagonal matrix") - E[g^2]. More...
mic::types::MatrixPtr< eT >	ED
	Decaying average of the squares of updates up to time t ("diagonal matrix") - E[delta Theta^2]. More...
mic::types::MatrixPtr< eT >	delta
	Calculated update. More...

Detailed Description

template<typename eT = float>
class mic::neural_nets::optimization::AdaDelta< eT >

Update using AdaDelta - adaptive gradient descent with running average E[g^2] and E[d^2].

Author

tkornuta

Definition at line 39 of file AdaDelta.hpp.

Constructor & Destructor Documentation

template<typename eT = float>

mic::neural_nets::optimization::AdaDelta< eT >::AdaDelta ( size_t  rows_, size_t  cols_, eT  decay_ = 0.9, eT  eps_ = 1e-8  )

inline

Constructor. Sets dimensions, values of decay (default=0.9) and eps (default=1e-8).

Parameters

rows_	Number of rows of the updated matrix/its gradient.
cols_	Number of columns of the updated matrix/its gradient.

Definition at line 47 of file AdaDelta.hpp.

References mic::neural_nets::optimization::AdaDelta< eT >::delta, mic::neural_nets::optimization::AdaDelta< eT >::ED, and mic::neural_nets::optimization::AdaDelta< eT >::EG.

Member Function Documentation

template<typename eT = float>

mic::types::MatrixPtr<eT> mic::neural_nets::optimization::AdaDelta< eT >::calculateUpdate ( mic::types::MatrixPtr< eT >  x_, mic::types::MatrixPtr< eT >  dx_, eT  learning_rate_  )

inlinevirtual

Calculates the update according to the AdaDelta update rule.

Parameters

x_	Pointer to the current matrix.
dx_	Pointer to current gradient of that matrix.
learning_rate_	Learning rate (default=0.001). NOT USED!

Implements mic::neural_nets::optimization::OptimizationFunction< eT >.

Definition at line 65 of file AdaDelta.hpp.

References mic::neural_nets::optimization::AdaDelta< eT >::decay, mic::neural_nets::optimization::AdaDelta< eT >::delta, mic::neural_nets::optimization::AdaDelta< eT >::ED, mic::neural_nets::optimization::AdaDelta< eT >::EG, and mic::neural_nets::optimization::AdaDelta< eT >::eps.

Member Data Documentation

template<typename eT = float>

eT mic::neural_nets::optimization::AdaDelta< eT >::decay

protected

Decay ratio, similar to momentum.

Definition at line 101 of file AdaDelta.hpp.

Referenced by mic::neural_nets::optimization::AdaDelta< eT >::calculateUpdate().

template<typename eT = float>

mic::types::MatrixPtr<eT> mic::neural_nets::optimization::AdaDelta< eT >::delta

protected

Calculated update.

Definition at line 113 of file AdaDelta.hpp.

Referenced by mic::neural_nets::optimization::AdaDelta< eT >::AdaDelta(), and mic::neural_nets::optimization::AdaDelta< eT >::calculateUpdate().

template<typename eT = float>

mic::types::MatrixPtr<eT> mic::neural_nets::optimization::AdaDelta< eT >::ED

protected

Decaying average of the squares of updates up to time t ("diagonal matrix") - E[delta Theta^2].

Definition at line 110 of file AdaDelta.hpp.

Referenced by mic::neural_nets::optimization::AdaDelta< eT >::AdaDelta(), and mic::neural_nets::optimization::AdaDelta< eT >::calculateUpdate().

template<typename eT = float>

mic::types::MatrixPtr<eT> mic::neural_nets::optimization::AdaDelta< eT >::EG

protected

Decaying average of the squares of gradients up to time t ("diagonal matrix") - E[g^2].

Definition at line 107 of file AdaDelta.hpp.

Referenced by mic::neural_nets::optimization::AdaDelta< eT >::AdaDelta(), and mic::neural_nets::optimization::AdaDelta< eT >::calculateUpdate().

template<typename eT = float>

eT mic::neural_nets::optimization::AdaDelta< eT >::eps

protected

Smoothing term that avoids division by zero.

Definition at line 104 of file AdaDelta.hpp.

Referenced by mic::neural_nets::optimization::AdaDelta< eT >::calculateUpdate().

---

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

• src/optimization/AdaDelta.hpp