libeblearn
|
#include <ebl_states.h>
Public Member Functions | |
gd_param () | |
Empty constructor, set all variables to zero. | |
gd_param (double leta, double ln, double l1, double l2, intg dtime, double iner, double a_v, intg a_p, double g_t) | |
Constructor. | |
Public Attributes | |
double | eta |
global step size | |
double | n |
stopping criterion threshold | |
double | decay_l1 |
L1 regularizer coefficient. | |
double | decay_l2 |
L2 regularizer coefficient. | |
intg | decay_time |
Time (in number of training samples) after which to decay values. | |
double | inertia |
momentum term | |
double | anneal_value |
annealing coefficient for the learning rate | |
intg | anneal_period |
Number of training samples beetween two annealings. | |
double | gradient_threshold |
threshold on square norm of gradient for stopping | |
int | niter_done |
for debugging purpose |
A class that contains all the parameters for a stochastic gradient descent update, including step sizes, regularizer coefficients...