ebl::supervised_trainer< Tdata, Tlabel > Class Template Reference

#include <EblTrainer.h>

List of all members.

Public Member Functions
	supervised_trainer (fc_ebm2< state_idx, int, state_idx > &m, parameter &p, ostream &cout=std::cout)
void	set_display (unsigned int nh, unsigned int nw, unsigned int h0=0, unsigned int w0=0, double zoom=1.0, int wid=-1, const char *title=NULL)
int	run (state_idx &input, infer_param &infp)
bool	test_sample (state_idx &input, int label, int &answer, infer_param &infp)
Idx< double >	learn_sample (state_idx &input, int label, gd_param &arg)
void	test (LabeledDataSource< Tdata, Tlabel > &ds, classifier_meter &log, infer_param &infp, bool display=false)
void	train (LabeledDataSource< Tdata, Tlabel > &ds, classifier_meter &log, gd_param &args, int niter)
void	compute_diaghessian (LabeledDataSource< Tdata, Tlabel > &ds, intg niter, double mu)
	compute hessian
void	resize_input (LabeledDataSource< Tdata, Tlabel > &ds)
Public Attributes
fc_ebm2< state_idx, int, state_idx > &	machine
parameter &	param
state_idx *	input
state_idx	energy
Idx< Tlabel >	label
intg	age
ostream &	cout

---

Detailed Description

template<class Tdata, class Tlabel>
class ebl::supervised_trainer< Tdata, Tlabel >

Generic Trainer Supervised Trainer

---

Member Function Documentation

template<class Tdata, class Tlabel>

int ebl::supervised_trainer< Tdata, Tlabel >::run	(	state_idx &	input,
		infer_param &	infp
	)			[inline]

take an input and a vector of possible labels (each of which is a vector, hence <label-set> is a matrix) and return the index of the label that minimizes the energy fill up the vector <energies> with the energy produced by each possible label. The first dimension of <label-set> must be equal to the dimension of <energies>.

template<class Tdata, class Tlabel>

bool ebl::supervised_trainer< Tdata, Tlabel >::test_sample	(	state_idx &	input,
		int	label,
		int &	answer,
		infer_param &	infp
	)			[inline]

Test a single sample and its label <label> (an integer). Returns true if the sample was correctly classified, false otherwise.

template<class Tdata, class Tlabel>

Idx< double > ebl::supervised_trainer< Tdata, Tlabel >::learn_sample	(	state_idx &	input,
		int	label,
		gd_param &	arg
	)			[inline]

perform a learning update on one sample. <sample> is the input sample, <label> is the desired category (an integer), <label-set> is a matrix where the i-th row is the desired output for the i-th category, and <update-args> is a list of arguments for the parameter update method (e.g. learning rate and weight decay).

template<class Tdata, class Tlabel>

void ebl::supervised_trainer< Tdata, Tlabel >::test	(	LabeledDataSource< Tdata, Tlabel > &	ds,
		classifier_meter &	log,
		infer_param &	infp,
		bool	display = false
	)			[inline]

Measure the average energy and classification error rate on a dataset. returns a list with average loss and proportion of errors

template<class Tdata, class Tlabel>

void ebl::supervised_trainer< Tdata, Tlabel >::train	(	LabeledDataSource< Tdata, Tlabel > &	ds,
		classifier_meter &	log,
		gd_param &	args,
		int	niter
	)			[inline]

train for <niter> sweeps over the training set. <samples> contains the inputs samples, and <labels> the corresponding desired categories <labels>. return the average energy computed on-the-fly. <update-args> is a list of arguments for the parameter update method (e.g. learning rate and weight decay).

template<class Tdata, class Tlabel>

void ebl::supervised_trainer< Tdata, Tlabel >::resize_input

(

LabeledDataSource< Tdata, Tlabel > &

ds

)

[inline]

Resize <input> based on the datasource. If <input> is not allocated, allocate it. TODO: If order or dimensions have changed, reallocate.

---

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

• /media/e/eblearn-trunk/libeblearn/include/EblTrainer.h
• /media/e/eblearn-trunk/libeblearn/include/EblTrainer.hpp