API_LowLvl::lib::mlp

Collaboration diagram for API_LowLvl::lib::mlp:

Data Structures
struct	SprMLP
	the main MLP structure More...

Macros
#define	spr_dt_nn_float_data
#define	SPR_MLP_MODIF_EVAL
#define	SPR_MLP_MODIF_TRAIN_X
#define	SPR_MLP_MODIF_TRAIN_GD
#define	SPR_MLP_MODIF_TRAIN_WGD
#define	SPR_MLP_MODIF_TRAIN_WGD2
#define	SPR_MLP_MODIF_TRAIN_SCG
#define	SPR_MLP_MODIF_TRAIN_WSCG
#define	SPR_MLP_MODIF_TRAIN_CG
#define	SPR_MLP_MODIF_TRAIN_WCG
#define	spr_mlp_check_training(mlp, flags)

Typedefs
typedef double	SprNNFloatCalc
typedef float	SprNNFloatData
typedef SprNNFloatCalc(*	SprMlpNorm )(SprNNFloatCalc dd, SprMLP *mlp)

Enumerations
enum	{ SPR_MLP_TM_ANY, SPR_MLP_TM_WGT, SPR_MLP_TM_WGTP, SPR_MLP_TM_CG }

Functions
void	mlp_bwd_init (SprMLP *mlp)
SprNNFloatData *	spr_mlp_eval (SprMLP *mlp, const float *x)
SprMLP *	spr_mlp_free (SprMLP *mlp)
int	spr_mlp_write (const char *fname, SprMLP *mlp, int write_dsc)
SprMLP *	spr_mlp_read (const char *fname)
int	spr_mlp_modif (SprMLP *mlp, int action,...)
SprNNFloatData *	spr_mlp_norm_output (SprMLP *mlp)
	Normalize the output (sum equals to 1.0). More...
SprNNFloatData *	mlp_train_fwd (SprMLP *mlp, float *x)
double	mlp_train_bwd (SprMLP *mlp)
SprNNFloatCalc	mlp_train_err_ls (SprMLP *mlp, const float *t, const float *wgt, SprNNFloatCalc eps)
SprNNFloatCalc	mlp_train_err_ml (SprMLP *mlp, int Ndx, SprNNFloatCalc eps)
SprNNFloatCalc	mlp_train_err_ml1 (SprMLP *mlp, int Ndx, SprNNFloatCalc eps)
int	mlp_update (SprMLP *mlp, int final)
int	mlp_update_WGDI (SprMLP *mlp)
void	spr_mlp_train_stats (SprStream *fd, SprMLP *mlp)
int	spr_mlp_train_state_io (const char *fname, SprMLP *mlp, int save)

Detailed Description

Macro Definition Documentation

#define spr_dt_nn_float_data

#define SPR_MLP_MODIF_EVAL

#define SPR_MLP_MODIF_TRAIN_X

#define SPR_MLP_MODIF_TRAIN_GD

#define SPR_MLP_MODIF_TRAIN_WGD

#define SPR_MLP_MODIF_TRAIN_WGD2

#define SPR_MLP_MODIF_TRAIN_SCG

#define SPR_MLP_MODIF_TRAIN_WSCG

#define SPR_MLP_MODIF_TRAIN_CG

#define SPR_MLP_MODIF_TRAIN_WCG

#define spr_mlp_check_training	(		mlp,
			flags
	)

Possible flags are: "MLP_TM_ANY" any traning method (not in eval/dummy train mode) "MLP_TM_WGT" W(GD/SCG/CG) training "MLP_TM_CG" (W)(S)CG training

Typedef Documentation

typedef double SprNNFloatCalc

typedef float SprNNFloatData

typedef SprNNFloatCalc(* SprMlpNorm)(SprNNFloatCalc dd, SprMLP *mlp)

Enumeration Type Documentation

anonymous enum

Enumerator
SPR_MLP_TM_ANY
SPR_MLP_TM_WGT
SPR_MLP_TM_WGTP
SPR_MLP_TM_CG

Function Documentation

void mlp_bwd_init

(

SprMLP *

mlp

)

Reset all buffers (prepare for evaluation) and set the input buffer.

SprNNFloatData* spr_mlp_eval	(	SprMLP *	mlp,
		const float *	x
	)

Perform all routines to evaluate/train the MLP for the given input.

SprMLP* spr_mlp_free

(

SprMLP *

mlp

)

int spr_mlp_write	(	const char *	fname,
		SprMLP *	mlp,
		int	write_dsc
	)

Write a (trained) MLP to disk. See mlp_read() for the conventions on the location of the accompanying parameters files. If the write_dsc flag is set, a new description file is written.

SprMLP* spr_mlp_read

(

const char *

fname

)

Read an MLP. The file fname contains the MLP network description. Parameter files are suposed to reside in the same subdirectory as the network description file.

int spr_mlp_modif	(	SprMLP *	mlp,
		int	action,
			...
	)

Modify some of the settings of the MLP. Return value: (-1) on error.
Possible actions: "MLP_MODIF_EVAL" Prepare the network for fast evaluation. "MLP_MODIF_TRAIN_X" Prepare the network for a dummy training session: evaluation + error measurement. "MLP_MODIF_TRAIN_GD" Prepare the network for training (gradient descent). "MLP_MODIF_TRAIN_WGD" Prepare the network for training with adaptive learning rates (each parameter has its own learning rate). Requires an extra float argument to specify the global learning rate. "MLP_MODIF_TRAIN_WGD2" Prepare the network for training with adaptive learning rates. Requires an extra float argument to specify the global learning rate. "MLP_MODIF_TRAIN_SCG" Prepare the network for training with the scaled conjugent gradient algorithm (Moller, M. Neural networks vol 6., pp 525-533, 1993). Requires one extra float arguments (sigma). "MLP_MODIF_TRAIN_WSCG" Prepare the network for training with the weighted scaled conjugent gradient algorithm. "MLP_MODIF_TRAIN_CG" Prepare the network for training with a conjugent gradient algorithm. Requires one extra float argument (sigma). "MLP_MODIF_TRAIN_WCG" Prepare the network for training with the weighted conjugent gradient algorithm. "MLP_MODIF_CLEAR" Clear training/fast-evaluation information. "MLP_MODIF_NORM" Define a new norm for the least-sqaures training. The arguments are a float representing the norm, and a pointer to a function that calculates that norm. If the function is NULL, a build in function is used.

SprNNFloatData* spr_mlp_norm_output

(

SprMLP *

mlp

)

Normalize the output (sum equals to 1.0).

SprNNFloatData* mlp_train_fwd	(	SprMLP *	mlp,
		float *	x
	)

Forward evaluation of the MLP for training. Returns the output of the MLP.

double mlp_train_bwd

(

SprMLP *

mlp

)

Do the backpropagation of the error, and return the regularisation error.

SprNNFloatCalc mlp_train_err_ls	(	SprMLP *	mlp,
		const float *	t,
		const float *	wgt,
		SprNNFloatCalc	eps
	)

Calculate the frame error (norm_e distance) and set the gradients for the output layer. The norm is set with the routine mlp_modif().
The weight vector wgt may be a NULL pointer.
Returns the frame error: sum(|t-o|^e).

SprNNFloatCalc mlp_train_err_ml	(	SprMLP *	mlp,
		int	Ndx,
		SprNNFloatCalc	eps
	)

Calculate the frame error (log likelihood) and set the gradients for the output layer. The gradient also incorporates a second order cost that forces the sum of the outputs to be close to 1.0.

Returns

the log likelihood of the frame, i.e. log(o[Ndx]/sum(o)).

SprNNFloatCalc mlp_train_err_ml1	(	SprMLP *	mlp,
		int	Ndx,
		SprNNFloatCalc	eps
	)

Calculate the frame error (log error) and set the gradients for the output layer. Returns the log error of the frame: E = sum(o[i])-log(o[Ndx])-1

int mlp_update	(	SprMLP *	mlp,
		int	final
	)

Perform all routines to evaluate/train the MLP for the given input.

int mlp_update_WGDI

(

SprMLP *

mlp

)

Update the adaptive learning rate (no update for the MLP-parameters). Returns 1 when enough data has been processed to estimate all parameters reliably.

void spr_mlp_train_stats	(	SprStream *	fd,
		SprMLP *	mlp
	)

Calculate and print statistics about the weights in the weighted training.

int spr_mlp_train_state_io	(	const char *	fname,
		SprMLP *	mlp,
		int	save
	)

Load/save the train status when using the weighted gradient learning scheme. Loading an empty file equals to resetting all information to their initial values.