mnist-neural-network

mnist-neural-network implements a neural network to recognize handwritten digits. I used MNIST database to train the network.

Installation and usage

We can use make utility to install the software:

$ cd implementation && make

after the installation, we can execute the software:

$ ./mnn_main --help

the helper has this form:

Usage: ./mnn_main [Options]

Options:
   --help                               Print this help
   --config-file <file>                 Specify config file to load
                                        (if you don't specify this option, "default_configuration.conf" will be loaded)
   --load-parameters <file>             Load weights and biases contained in the file
                                        (you must use '--save-parameters' option in a previous execution)
   --save-parameters <file>             Save weights and biases in the file
   --image <file>                       Predict number of the image
                                        (file's name and its content must respect some constrains)
   --train                              Train the neural network
                                        (if you use "--save-parameters", this option will be implicit)
   --show-image <file>                  Show specified image
   --export-parameters <file>           Save weights and biases as matrices and arrays in the file with Javascript syntax

Config file

Before executing the neural network, we can modify the config file. It contains the hyperparameters of the network. The default configuration looks like:

{
    hidden-layers: 20;
    epochs: 30;
    learning-rate: 3;
    mini-batches: 10;
};

implementation/default_configuration.conf file contains all information about parameters and the values that they can assume (for example hidden-layers could be an array).

Execution

This video illustrates a typical execution of the software. Since training can require several times to terminate every epoch, we assume that we had already trained the network using --save-parameters option for saving the parameters inside params_1.txt. Moreover, this option will execute the training automatically as we had specified in the helper:

$ ./mnn_main --save-parameters data/output/params_1.txt

is equal to:

$ ./mnn_main --save-parameters data/output/params_1.txt --train

Tools

I created a set of tools to handle images. All of them, except extract_test_information_from_mnist_db.c, require OpenCV library.

Other information

• if you modify the configuration file and, for some reason, it will be incorrect, you can enable debug uncommenting line 16 at mnn_parser_config_file.h file to check the error;
• if you want to speed up the training, you can decrease the number of training and number of tests (at the end of each epoch, the program will check the number of correct prediction on 10.000 images). It's enough to modify constants N_TRAINING_IMAGES and N_TEST_IMAGES (file mnn_config.h line 11 and line 14).