darwinAPI

Made by Jean-Adrien DUCASTAING (ducast_j) for the OCR project during the third semester at EPITA.

NOTE:

This API is a quite unprofessional noob-friendly FFN tool. As the result of my first C project and also my first implementation of a neural network, it can be quite messy and don't reflect a professional code and architecture quality.
This implementation use a full-matrix approach of the FFN using sigmoid activation functions and vanilla gradient descent without any other optimization.

API's Matrix System:

darwinAPI use it's own built-in matrix library using a 1-D array for basics optimization purpose.
Here's a simple example how to make a 3x2 matrix:

double data[6] = {11, 12, 21, 22, 31, 32};
Matrix myMatrix = createMatrix(&data, 3, 2);

Note that darwinAPI is a full-matrix FFN implementation, all the elements that you are manipulating are matrix, for instance the input matrix n x p will correspond to a data set of n inputs of p elements. Of course you can input as many elements as you want since the only requirement is that the number of column match the size of the input layer.

FFN Functions:

-Create a new FFN: FFNet createFFN(int layersSizes[], layersNb)
-Train a FFN: void trainFFNet(FFNet* myFFNet, int epochs, Matrix* inputs, Matrix* desired_outputs)
-Compute with a FFN: Matrix forwardPropagation(FFNet* myFFNet, Matrix inputs)
-Free a FFN: void freeFFNet(FFNet* myFFNet)
-Save a FFN to a file: void serialFFNet(FFNet* myFFNet, char* myfilename)
-Load a FFN from a file FFNet loadFFNet(char* myfilename)

Quick Start (demo with a XOR):

Let's make quickly a 2-3-1 FFN capable of learning a XOR.

#include "matrix.h"
#include "activ.h"
#include "ffnet.h"
#include "backprop.h"
#include "serialize.h"
#include "ffnet.h"
#include <stdlib.h>
#include <time.h>
#include <stdio.h>
#include <math.h>

int main(){
    int layers[3] = {2,3,1};
    FFNet = myFFNet(layers, 3);
    
    double inputs[8] = {0, 0, 0, 1, 1, 0, 1, 1};
    Matrix inputsMat = createMatrix(inputs, 4, 2);
    
    double desired_outputs = {0, 1, 1, 0};
    Matrix desired_outputsMat = createMatrix(desired_outputs, 4, 1);
    
    trainFFNet(&myFFNet, 250000, &inputsMat, &desired_ouputsMat);
    Matrix output = forwardPropagation(&myFFNet, inputsMat);
    
    saveFFNet(&myFFNet, "xor_demo");
    free(&myFFNet);
}