train_main.c File Reference

Entry point for the CNN training binary. More...

#include "src/cnn/cnn.h"
#include "src/cnn/dataset.h"
#include "src/cnn/model.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <time.h>
#include <sys/stat.h>

Include dependency graph for train_main.c:

Classes
struct	TrainArgs
	Parsed command-line options for the train binary. More...

Macros
#define	DEFAULT_DATA_DIR   "training_data/"
#define	DEFAULT_MODEL_DIR   "models/"
#define	MAX_EPOCHS   50
#define	VAL_SPLIT   0.15f /* fraction of data kept for validation */
#define	ES_PATIENCE   5 /* epochs without improvement before stopping */
#define	ES_MIN_DELTA   1e-4f /* minimum val-loss improvement to count */

Functions
static void	usage (const char *prog)
	Print usage information to stderr.
static int	parse_args (int argc, char **argv, TrainArgs *args)
	Parse argv into a TrainArgs structure.
static double	eval_loss (CNN *net, const Sample *samples, size_t n, double *out_acc)
	Compute loss and accuracy on a slice of samples (no gradient update).
static void	train_loop (CNN *net, Dataset *ds, int n_epochs)
	Run the training loop with early stopping.
int	main (int argc, char **argv)

Detailed Description

Entry point for the CNN training binary.

Usage:

./train [--data <dir>] [--output <file>] [-j<N>]

Options: –data <dir> Training data root directory (default: training_data/). –output <file> Output model filename (default: models/model_<ts>.bin). -j<N> Number of loader threads (default: 1).

Example:

./train --data training_data/ --output my_model.bin -j4

Exit codes: 0 Success — model saved to the output path. 1 Argument error. 2 Dataset loading error. 3 Model save error.

Macro Definition Documentation

DEFAULT_DATA_DIR

#define DEFAULT_DATA_DIR   "training_data/"

DEFAULT_MODEL_DIR

#define DEFAULT_MODEL_DIR   "models/"

ES_MIN_DELTA

#define ES_MIN_DELTA   1e-4f /* minimum val-loss improvement to count */

ES_PATIENCE

#define ES_PATIENCE   5 /* epochs without improvement before stopping */

MAX_EPOCHS

#define MAX_EPOCHS   50

VAL_SPLIT

#define VAL_SPLIT   0.15f /* fraction of data kept for validation */

Function Documentation

eval_loss()

double eval_loss ( CNN * net, const Sample * samples, size_t n, double * out_acc )

static

Compute loss and accuracy on a slice of samples (no gradient update).

Parameters

net	Trained CNN.
samples	Pointer to first sample.
n	Number of samples.
out_acc	Filled with accuracy in [0, 1].

Returns

Average cross-entropy loss.

Here is the call graph for this function:

Here is the caller graph for this function:

main()

int main	(	int	argc,
		char **	argv )

Here is the call graph for this function:

parse_args()

int parse_args ( int argc, char ** argv, TrainArgs * args )

static

Parse argv into a TrainArgs structure.

Parameters

argc	Argument count.
argv	Argument vector.
args	Output structure (caller-allocated).

Returns

0 on success, -1 on parse error.

Here is the caller graph for this function:

train_loop()

void train_loop ( CNN * net, Dataset * ds, int n_epochs )

static

Run the training loop with early stopping.

The dataset is split into a training portion (1 - VAL_SPLIT) and a validation portion (VAL_SPLIT) after an initial shuffle.

Early stopping: training halts when validation loss has not improved by more than ES_MIN_DELTA for ES_PATIENCE consecutive epochs. The best weights observed so far are restored before returning.

Parameters

net	Initialised CNN.
ds	Full dataset (will be shuffled in-place).
n_epochs	Maximum number of epochs.

Here is the call graph for this function:

Here is the caller graph for this function:

usage()

void usage ( const char * prog )

static

Print usage information to stderr.

Parameters

prog	Program name (argv[0]).

Here is the caller graph for this function: