mini_batch_gd.c File Reference

Mini-batch stochastic gradient descent training example using DMLFW. More...

#include <dmlfw.h>
#include <unistd.h>
#include <stdio.h>
#include <stdlib.h>

Include dependency graph for mini_batch_gd.c:

dmlfw_mat_double *	buffer_matrix_1 = NULL
dmlfw_mat_double *	buffer_matrix_1_shuffled = NULL
dmlfw_mat_double *	buffer_matrix_2 = NULL
dmlfw_mat_double *	buffer_matrix_2_shuffled = NULL
uint32_t	BUFFER_SIZE = 50
void	data_loader (void *x, void *y, uint64_t from_row, uint32_t how_many_rows)
	Supplies mini-batch data to gradient descent algorithm.
#define	FREQUENCY_OF_PRINTING_COST   100
void	get_from_file_buffer (dmlfw_mat_double **target_matrix, uint64_t from_row, uint32_t how_many_rows)
	Retrieves a data block for training from buffers.
dmlfw_gradient_descent_options *	get_gradient_descent_options ()
	Creates and configures options for mini-batch stochastic gradient descent.
FILE *	gnuplot
void	init_buffers ()
	Initializes minibatch data buffers before training.
#define	LEARNING_RATE   0.00001
int	main ()
	Main entry point running the mini-batch stochastic gradient descent training example.
uint32_t	MINI_BATCH_SIZE = 46
#define	MODEL_FILE_NAME   "example-3-model.csv"
dmlfw_column_vec_double *	model_squared_sum_vector = NULL
dmlfw_row_vec_double *	model_transposed = NULL
uint64_t	number_of_columns_in_training_examples = 0
#define	NUMBER_OF_ITERATIONS   100000
uint64_t	number_of_training_examples = 0
int	on_iteration_complete (uint64_t iteration_number, void *y, void *predicted_y, void *model, double regularization_parameter)
	Callback after each gradient descent iteration to log cost and plot progress.
dmlfw_column_vec_double *	prediction_error = NULL
dmlfw_row_vec_double *	prediction_error_transposed = NULL
void	print_error_and_exit ()
	Prints ml-framework error string and terminates program.
dmlfw_column_vec_double *	product_vector = NULL
#define	REGULARIZATION_PARAMETER   0.5
#define	SHOW_GRAPH   1
#define	TRAINING_DATASET   "IceCreamSales_training_examples.csv"
dmlfw_mat_double *	x_matrix = NULL
dmlfw_mat_double *	xy_matrix = NULL
dmlfw_column_vec_double *	y_vector = NULL

Detailed Description

Mini-batch stochastic gradient descent training example using DMLFW.

Author

Mohammed Daniyal

Version

1.0

Date

2025-09-27

This program trains a linear regression model on the IceCreamSales dataset using mini-batch stochastic gradient descent with regularization. It manages data buffering, shuffling, and mini-batch provision for efficient training.

The training progress is logged and optionally visualized with gnuplot showing cost descent and fitted line graphs. The final model parameters are saved to a CSV file.

Global buffers are used for dataset handling, and a custom progress callback logs iteration-wise cost with early termination support.

Usage: ./mini_batch_gd

Macro Definition Documentation

FREQUENCY_OF_PRINTING_COST

#define FREQUENCY_OF_PRINTING_COST   100

Frequency of printing/logging cost

LEARNING_RATE

#define LEARNING_RATE   0.00001

Learning rate for gradient descent

MODEL_FILE_NAME

#define MODEL_FILE_NAME   "example-3-model.csv"

Output model CSV file path

NUMBER_OF_ITERATIONS

#define NUMBER_OF_ITERATIONS   100000

Maximum number of gradient descent iterations

REGULARIZATION_PARAMETER

#define REGULARIZATION_PARAMETER   0.5

Regularization parameter lambda

SHOW_GRAPH

#define SHOW_GRAPH   1

Enable or disable plotting with gnuplot (1 = enabled)

TRAINING_DATASET

#define TRAINING_DATASET   "IceCreamSales_training_examples.csv"

Training dataset CSV file path

Function Documentation

data_loader()

void data_loader	(	void *	x,
		void *	y,
		uint64_t	from_row,
		uint32_t	how_many_rows
	)

Supplies mini-batch data to gradient descent algorithm.

Parameters

x	Pointer to feature matrix pointer.
y	Pointer to target vector pointer.
from_row	Starting row of data to load.
how_many_rows	Number of rows to load.

get_from_file_buffer()

void get_from_file_buffer	(	dmlfw_mat_double **	target_matrix,
		uint64_t	from_row,
		uint32_t	how_many_rows
	)

Retrieves a data block for training from buffers.

Parameters

target_matrix	Matrix pointer to fill.
from_row	Starting row index in dataset.
how_many_rows	Number of rows to retrieve.

get_gradient_descent_options()

dmlfw_gradient_descent_options * get_gradient_descent_options

(

)

Creates and configures options for mini-batch stochastic gradient descent.

Returns

Pointer to configured options struct or NULL on error.

init_buffers()

void init_buffers

(

)

Initializes minibatch data buffers before training.

main()

int main

(

)

Main entry point running the mini-batch stochastic gradient descent training example.

Returns

0 on success or error code.

on_iteration_complete()

int on_iteration_complete	(	uint64_t	iteration_number,
		void *	y,
		void *	predicted_y,
		void *	model,
		double	regularization_parameter
	)

Callback after each gradient descent iteration to log cost and plot progress.

Parameters

iteration_number	Current iteration index.
y	Actual target values.
predicted_y	Model predicted values.
model	Current model parameters.
regularization_parameter	Lambda regularization value.

Returns

0 to continue optimization, -1 to terminate early.

print_error_and_exit()

void print_error_and_exit

(

)

Prints ml-framework error string and terminates program.

Variable Documentation

buffer_matrix_1

dmlfw_mat_double* buffer_matrix_1 = NULL

Primary buffer matrix holding dataset block for training

buffer_matrix_1_shuffled

dmlfw_mat_double* buffer_matrix_1_shuffled = NULL

Shuffled copy of primary buffer matrix

buffer_matrix_2

dmlfw_mat_double* buffer_matrix_2 = NULL

Secondary buffer matrix holding remaining dataset block

buffer_matrix_2_shuffled

dmlfw_mat_double* buffer_matrix_2_shuffled = NULL

Shuffled copy of secondary buffer matrix

BUFFER_SIZE

uint32_t BUFFER_SIZE = 50

Buffer size, adjustable at runtime for batch loading

gnuplot

FILE* gnuplot

Gnuplot pipe pointer used for plotting graphs

MINI_BATCH_SIZE

uint32_t MINI_BATCH_SIZE = 46

Mini batch size, adjustable at runtime

model_squared_sum_vector

dmlfw_column_vec_double* model_squared_sum_vector = NULL

Vector holding squared sum of model parameters for regularization calculation

model_transposed

dmlfw_row_vec_double* model_transposed = NULL

Transposed vector of current model parameters during callback

number_of_columns_in_training_examples

uint64_t number_of_columns_in_training_examples = 0

Number of columns (features + target) in training dataset

number_of_training_examples

uint64_t number_of_training_examples = 0

Total number of training examples (rows) in dataset

prediction_error

dmlfw_column_vec_double* prediction_error = NULL

Vector holding prediction errors during training progress callback

prediction_error_transposed

dmlfw_row_vec_double* prediction_error_transposed = NULL

Transposed version of prediction_error vector

product_vector

dmlfw_column_vec_double* product_vector = NULL

Temporary vector used for intermediate matrix products

x_matrix

dmlfw_mat_double* x_matrix = NULL

Feature matrix pointer used in data loading for mini-batches

xy_matrix

dmlfw_mat_double* xy_matrix = NULL

Combined feature-target matrix used temporarily for data loading

y_vector

dmlfw_column_vec_double* y_vector = NULL

Target vector pointer used in data loading for mini-batches