Treinamento da Rede Neural

nn_train.py

def nnTrain(epsilon, alpha, max_iter):
    input_layer_size = x_train.shape[1]
    hidden_layer_size = 800
    num_labels = 10
    theta_1, theta_2 = randomInit(input_layer_size, hidden_layer_size, num_labels)
    
    for i in range(max_iter):
      J_theta, Theta1_grad, Theta2_grad = nnRegCostFunction(
          theta_1, theta_2, x_train, y_train,
          input_layer_size, hidden_layer_size, num_labels)

      theta_1 = theta_1 - alpha * Theta1_grad
      theta_2 = theta_2 - alpha * Theta2_grad
      
      acc = accuracy(theta_1, theta_2, x_train, y_train, y_train_classif)
      print(f'Iteração {i+1:2d}:  custo={J_theta:5.2f}  acurácia={100*acc:4.1f}%')
      
      if 1-acc < epsilon:
        break
      
    return theta_1, theta_2