NN

Neural Network with back propagation learning algorithm written in C.

Usage

The project is intended to be used as a library.
However, there is a demo program examples/digits.c included, which uses
the MNIST database
to train itself and classify images of handwritten digits.

git submodule update --init --recursive
cd build
cmake ..
make

To run the example:

./digits five.png

The program will perform the training on it’s first run.

Quickstart

Please check include/nn/nn.h for better API explanation.

/* Allocates a new (empty) network with n inputs. */
neuralnetwork *nn_create(int n);
/* Deallocates the network and all allocated data. */
void nn_destroy(neuralnetwork *nn);
/* Adds a new layer to the network with specified number of neurons.
 * Weights can be passed as a matrix stored in a double array.
 * NULL initializes weights randomly and biases with 0.
 * Possible values for activations:
 * IDENTITY, STEP, TANH, RELU, RELU_LEAKY, GAUSSIAN, SIGMOID, SOFTPLUS
 * Xavier initialization is used for all activation functions, except RELUs,
 * for which Kaiming is used. */
void nn_addlayer(neuralnetwork *nn, int nodes, double *weights, double *biases, int activation);
/* Forward propagates a given input through the network.
 * Returns pointer to the output array. */
double *nn_forwardpropagate(neuralnetwork *nn, double *input);
/* Performs forward propagation followed by the backpropagation to teach the network.
 * Passing learning rate of 0 will not perform back propagation.
 * Returns mean squared error of the forward pass. */
double nn_backpropagate(neuralnetwork *nn, double *input, double *target, double learningrate);
/* Useful functions for getting the number of inputs/outputs of the network. */
int nn_ninputs(neuralnetwork *nn);
int nn_noutputs(neuralnetwork *nn);
/* File I/O functions. Store/read the network as a binary file to keep the double precision.
 * 0 or NULL is returned on failure. */
int nn_writefile(const neuralnetwork *nn, const char *filename);
neuralnetwork *nn_readfile(const char *filename);

License

This project is licensed under the MIT License - see the LICENSE.md file for details