neuronal.py

import math import numpy as np # Fonction de trasfert tahn, et sa dérivée. def sigmoid(x): return np.tanh(x) def dsigmoid(x): return 1.0 - x**2 class Reseau_Neurones(object): def __init__(self, *args): # On fixe la structure du RN (ie. nb et taille des couches) # self.layers : prendra les valeurs de la propagation de X dans le RN n = len(args) self.layers = [np.ones(args[i] + (i==0)) for i in range(0, n)] # liste des matrices Wn, à valeurs rand entre -0.2 et 0.2 self.weights = list() for i in range(n-1): R = np.random.random((self.layers[i].size, self.layers[i+1].size)) self.weights.append((2*R-1)*0.20) # Qu'es-ce ? self.m = [0 for i in range(len(self.weights))] # Fonction de propagation de l'entrèe dans le RN def propag(self, inputs): # on donne l'entrée X au RN, en gardant un "1" à la fin (seuil) self.layers[0][:-1] = inputs for i in range(1, len(self.layers)): # sigmoide du produit matriciel X.Wn self.layers[i] = sigmoid(np.dot(self.layers[i-1], self.weights[i-1])) return self.layers[-1] def retroPropag(self, inputs, outputs, a=0.1, m=0.1): error = outputs - self.propag(inputs) de = error*dsigmoid(self.layers[-1]) deltas = list() deltas.append(de) for i in range(len(self.layers)-2, 0, -1): deh = np.dot(deltas[-1], self.weights[i].T) * dsigmoid(self.layers[i]) deltas.append(deh) deltas.reverse() for i, j in enumerate(self.weights): layer = np.atleast_2d(self.layers[i]) delta = np.atleast_2d(deltas[i]) dw = np.dot(layer.T,delta) self.weights[i] += a*dw + m*self.m[i] self.m[i] = dw RN = Reseau_Neurones(2, 3, 1) X = np.array(([3,5], [5,1], [10,2]), dtype = float) y = np.array(([75], [82], [93]), dtype = float) pat = (((0, 0), 0), ((0, 1), 1), ((1, 0), 1), ((1, 1), 0)) n = RN(2, 3, 1) for i in range(1000): for p in pat: n.backPropagate(p[0], p[1]) for p in pat: print n.update(p[0])