4NN_4.py

def train(x,y,w1,w2,w3,b1,b2,b3) :

    # feed forward
    inp_l = x
    l1 = sigmoid(np.dot(inp_l,w1) + b1)
    l2 = sigmoid(np.dot(l1,w2) + b2)
    l3 = sigmoid(np.dot(l2,w3) + b3)

    #backpropagation
    delta_l3 = (y - l3)*sigmoid_prime(l3)          
    delta_l2 = delta_l3.dot(w3.T)*sigmoid_prime(l2) 
    delta_l1 = delta_l2.dot(w2.T)*sigmoid_prime(l1) 

  
    #update weights
    w3 += (np.dot(l2.T,delta_l3))
    w2 += (np.dot(l1.T,delta_l2))
    w1 += (np.dot(inp_l.T,delta_l1))

    #  update biases
    b3 = b3 + (np.sum(delta_l3,axis=0))
    b2 = b2 + (np.sum(delta_l2,axis=0))
    b1 = b1 + (np.sum(delta_l1,axis=0))
    #print(b1.shape)

    return w1,w2,w3,b1,b2,b3