Collaboration diagram for API_HighLvl::lib::nn:

Data Structures
struct	SprNN

Typedefs
typedef struct SprNNStreamO_t	SprNNStreamO

typedef struct SprNNStreamC_t	SprNNStreamC
	close an open stream More...

typedef struct SprNNIOEl_t	SprNNIOEl

typedef struct SprNNI_t	SprNNI
	NN interface. More...

typedef void (	SprNNDataIn )(void restrict layer_val, void restrict src, int Nel)
	function point to handle the input data More...

typedef void (	SprNNDataOut )(void restrict dst, const void restrict layer_val, int Nel)
	function point to handle the output data More...

Enumerations
enum	{ SPR_NN_EVAL, SPR_NN_TRAIN, SPR_NN_NOVECTOR, SPR_NN_NOSCALAR, SPR_NN_PARALLEL, SPR_NN_ASYNC, SPR_NN_TRAIN2, SPR_NN_NOWARN }

Functions
int	spr_nn_dump (SprStream fd, const SprNN restrict nn)

SprNN *	spr_nn_free (SprNN *nn)

SprNN *	spr_nn_init (SprStream fd, const char fname, int flags, SprVarlist *vars)

void *	spr_nn_data_in_memcpy (void restrict layer_val, void restrict src, int Nel)
	a standard implementation to handle the input data More...

void *	spr_nn_data_out_memcpy (void restrict dst, const void restrict layer_val, int Nel)
	a standard implementation to handle the output data More...

void *	spr_nn_data_out_null (void restrict dst, const void restrict layer_val, int Nel)
	ignore the output data (flush the system) More...

int	spr_nni_feed_input (SprNNI restrict nni, void restrict data, SprNNDataIn func_get)

int	spr_nni_read_output (SprNNI restrict nni, int block, void restrict data, SprNNDataOut func_put)

SprNNI *	spr_nni_free (SprNNI *restrict nni)

SprNNI *	spr_nn_interface (SprNN *restrict nn, int flags, int MT)

Detailed Description

Typedef Documentation

typedef struct SprNNStreamO_t SprNNStreamO

open a file for streaming the data to the read functions or for writing back the trained parameters

typedef struct SprNNStreamC_t SprNNStreamC

close an open stream

typedef struct SprNNIOEl_t SprNNIOEl

connect and stream open/close elements are stored in one IO list

typedef struct SprNNI_t SprNNI

NN interface.

typedef void*(* SprNNDataIn)(void *restrict layer_val,void *restrict src,int Nel)

function point to handle the input data

typedef void*(* SprNNDataOut)(void *restrict dst,const void *restrict layer_val,int Nel)

function point to handle the output data

Enumeration Type Documentation

anonymous enum

Enumerator
SPR_NN_EVAL	read the data with the purpose to evaluate the NN (can be combined with TRAIN)
SPR_NN_TRAIN	read the data with the purpose to train the NN (can be combined with EVAL)
SPR_NN_NOVECTOR	do not use vector operations (require scalar operations)
SPR_NN_NOSCALAR	do not use scalar operations (require vector operations)
SPR_NN_PARALLEL	process multiple frames in parallel (reduces strain on memory resulting in approx. 2x faster processing)
SPR_NN_ASYNC	allow asynchronuous training updates, i.e. the parameters are updated while (instead of before) the gradients for the next batch are computed.
SPR_NN_TRAIN2	estimate second order statistics as well (learning rate per parameter)
SPR_NN_NOWARN	do not warn on slow/inefficient constructs

Function Documentation

int spr_nn_dump	(	SprStream *	fd,
		const SprNN *restrict	nn
	)

SprNN* spr_nn_free ( SprNN * nn )

SprNN* spr_nn_init	(	SprStream *	fd,
		const char *	fname,
		int	flags,
		SprVarlist *	vars
	)

Read a configuration from an open stream fd, or from the file with name fname (if fd equals NULL), or using spr_ssp_next_line() using fd as 'line_array' (if the bit flag SPR_NN_SSP_READ is set). The configuration defined the topology (number of layers, connections, ...) and the sources (files, ...) from which the parameters must be read. Before anything can be done with the NN, an interface has to be opened. The flags flags (and optionally some extra flags in string format) can limit the type of interfaces (vector, scalar, eval, train, ...) that can be opened, which allows more agressive fine-tuning of the parameter storage. The argument vars can be used to pass existing variables to this routine, e.g. specifying the size of the NN. Use NULL to indicate that no exiting variables are to be passed to the NN initialisation.

Returns: a handler to the configuration on success and NULL on failure.

Note: the function can also be called to print information to fd concerning the NN config file format by specifying (-1) as flags flags

void* spr_nn_data_in_memcpy	(	void *restrict	layer_val,
		void *restrict	src,
		int	Nel
	)

a standard implementation to handle the input data

Copy Nel floating point values from src to layer_val.

Returns: The updated src pointer (src+Nel*sizeof(SprNNFltS)) on success and NULL on error.

Parameters

layer_val	destination for the requested data (node values of some input layer)
src	the source pointer, points to where in memory the input data can be read from
Nel	number of elements requested

void* spr_nn_data_out_memcpy	(	void *restrict	dst,
		const void *restrict	layer_val,
		int	Nel
	)

a standard implementation to handle the output data

Copy Nel floating point values from layer_val to dst.

Returns: The updated dst pointer (dst+Nel*sizeof(SprNNFltS)) on success and NULL on error.

Parameters

dst	the destination pointer, points to where in memory the output data must be written to
layer_val	the data (node values of some input layer) that must be written
Nel	number of elements to write

void* spr_nn_data_out_null	(	void *restrict	dst,
		const void *restrict	layer_val,
		int	Nel
	)

ignore the output data (flush the system)

Ignore the output data.

Returns: a valid non-NULL pointer (layer_val+Nel).

int spr_nni_feed_input	(	SprNNI *restrict	nni,
		void *restrict	data,
		SprNNDataIn	func_get
	)

Copy the input data referred to be data to the input(s) of the NN and execute the 'fwd' routine in nni if a full set of frames is collected (SPR_NN_N_PARALLEL in the case of parallel evaluation of frames, 1 frame otherwise). The function func_get known how to handle the data pointer (which may be a memory pointer, a stream, ...). If data equals NULL, any pending non full set of frames is closed and processed by the 'fwd' routine in nni.

Returns: (-1) on failure (thread error or error with the input data), otherwise the number of frames send to the 'eval' routine in nni is returned (can be zero if parallel execution is requestioned and not enough frame have been collected).

Parameters

nni	NN interface
data	the input data (func_get known how to get the data)
func_get	function that get/copies the input data to the input layers

int spr_nni_read_output	(	SprNNI *restrict	nni,
		int	block,
		void *restrict	data,
		SprNNDataOut	func_put
	)

Check if output is available (block==0), and if no output is available either wait till output is available (block!=0) or until at least one input slot is available. If output is available, copy the next frame of the NN to wherever data points (memory, stream, ...) using func_put.

Returns: 0 if no output data is available, 1 if output data is available (and copied), and (-1) on failure (thread error, copying the output data failed).

Parameters

nni	NN interface
block	wait till output is available
data	where to write the output data (func_put known how to write the data)
func_put	function that put/copies the output layer data to wherever data points

SprNNI* spr_nni_free ( SprNNI *restrict nni )

Free the resources (memory, threads) associated with the NN interface nni.

Parameters

nni	NN interface

SprNNI* spr_nn_interface	(	SprNN *restrict	nn,
		int	flags,
		int	MT
	)

Make an interface toward the NN for evaluation (bit flag SPR_NN_EVAL set in <flags) and/or for training (bit flag SPR_NN_TRAIN set in flags). Configure for MT worker threads (use MT==0 for single threaded use).

Other relevant bit flags are:

SPR_NN_NOVECTOR: do not use vector operations (require scalar operations)
SPR_NN_NOSCALAR: do not use scalar operations (require vector operations)
SPR_NN_PARALLEL: process multiple frames in parallel (reduces strain on memory resulting in approx. 2x faster processing)
SPR_NN_ASYNC: allow asynchronuous training updates, i.e. the parameters are updated while (instead of before) the gradients for the next batch are computed
SPR_NN_NOWARN: do not warn on slow/inefficient constructs

Returns: the interface of NULL on error.

Data Structures

Typedefs

Enumerations

Functions

Detailed Description

Typedef Documentation

Enumeration Type Documentation

Function Documentation