ann.py

import numpy as np

def sigmoid(num):
    return 1 / (1 + np.exp(-num))

def deriv_sigmoid(z):
    return z * (1 - z)

class ANN:
    def __init__(self, x, y):
        self.input = x 
        self.output = np.zeros(y.shape)
        self.y = y
        self.weight1 = np.random.uniform(low = -0.5, high = 0.5 , size = (self.input.shape[1],500))
        # self.b1 = np.random.uniform(low = -0.5, high = 0.5 , size = (1,100))
        self.weight2 = np.random.uniform(low = -0.5, high = 0.5 , size = (500,500))
        # self.b2 = np.random.uniform(low = -0.5, high = 0.5 , size = (1,100))
        self.weight3 = np.random.uniform(low = -0.5, high = 0.5 , size = (500,1))
        # self.b3 = np.random.uniform(low = -0.5, high = 0.5 , size = (1,1))
        self.learningRate = 0.0001

    def forward(self):
        self.layer1 = sigmoid(np.dot(self.input, self.weight1) )
        self.layer2 = sigmoid(np.dot(self.layer1, self.weight2) )
        self.output = sigmoid(np.dot(self.layer2,self.weight3) )
    
    def back(self):
        deriv_weights3 = np.dot(self.layer2.T, (2*(self.y - self.output) * deriv_sigmoid(self.output)))
        deriv_weights2 = np.dot(self.layer1.T, (np.dot(2*(self.y - self.output) * deriv_sigmoid(self.output), self.weight3.T) * deriv_sigmoid(self.layer2)))
        deriv_weights1 = np.dot(self.input.T, (np.dot(2*(self.y - self.output) * deriv_sigmoid(self.layer2), self.weight2.T) * deriv_sigmoid(self.layer1)))

        self.weight1 += deriv_weights1 * self.learningRate
        self.weight2 += deriv_weights2 * self.learningRate
        self.weight3 += deriv_weights3 * self.learningRate

        # self.b1 += deriv_weights1.sum() * self.learningRate
        # self.b2 += deriv_weights2.sum() * self.learningRate
        # self.b3 += deriv_weights3.sum() * self.learningRate