DimaNet

DimaNet is a small neural network library written in C11. It is a fork of GENANN. The whole thing is two files: dimanet.c and dimanet.h. No dependencies, just a C compiler and libm.

Features

• Feedforward networks with any number of hidden layers
• Activation functions: sigmoid, tanh, ReLU, leaky ReLU, ELU, Softplus, SiLU, GELU and more
• Initializers: uniform, normal, Xavier, He, LeCun
• Backpropagation with optional momentum and gradient clipping
• Seeded randomness so runs are reproducible
• Save and load networks as text files

Build

With make:

make            # builds libdimanet.a and libdimanet.so in build/
make test       # runs the test suite
make examples   # builds the examples in build/
make install    # installs to /usr/local

With cmake:

cmake -B build
cmake --build build
ctest --test-dir build

Quick start

#include "dimanet.h"

int main(void) {
    dimanet_seed(1);

    /* 2 inputs, 1 hidden layer with 3 neurons, 1 output. */
    dimanet *ann = dimanet_init(2, 1, 3, 1);

    const double input[2] = {0.0, 1.0};
    const double target[1] = {1.0};

    dimanet_train(ann, input, target, 0.1);
    printf("%f\n", dimanet_run(ann, input)[0]);

    dimanet_free(ann);
    return 0;
}

Compile with cc example.c dimanet.c -lm.

Examples

The examples/ directory has five programs, including XOR learning and the IRIS classifier. See examples/README.md.

License

GPL-3.0. Derived from GENANN (MIT), copyright (c) 2015-2016 Lewis Van Winkle.