Wrapper to represent the parameters used during learning. More...

#include <crf.h>

Public Attributes
double	c2 = 1
	The regularization parameter.

double	delta = 1e-5
	The convergence threshold. More...

uint64_t	period = 10
	The period used to check for convergence.

double	lambda = 0
	The transformed regularization parameter. More...

double	t0 = 0
	The offset for the learning rate. More...

uint64_t	max_iters = 1000
	The maximum number of iterations to allow the gradient descent to run for.

double	calibration_eta = 0.1
	The initial starting value for \(\eta\), the learning rate, during calibration.

double	calibration_rate = 2.0
	The rate at which to adjust \(\eta\) during calibration.

uint64_t	calibration_samples = 1000
	The maximum number of samples to use during calibration.

uint64_t	calibration_trials = 10
	The maximum number of candidate \(\eta\)-s to consider during calibration.

Detailed Description

Wrapper to represent the parameters used during learning.

The defaults are sane, and so most users should simply initialize the default parameter object when training the crf.

Member Data Documentation

double meta::sequence::crf::parameters::delta = 1e-5

The convergence threshold.

Once the difference in the loss between period iterations is less than this value, learning will stop.

double meta::sequence::crf::parameters::lambda = 0

The transformed regularization parameter.

(This is set by the CRF internally based on c2 and the training set size.)

double meta::sequence::crf::parameters::t0 = 0

The offset for the learning rate.

The learning rate follows a the following schedule:

\(\eta = \frac{1}{\lambda * (t_0 + t)}\)

where \(t\) is the number of examples seen thus far.

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