带有方程学习器网络的非线性回归

编程问答 2022-05-24

问题描述

我正在处理 this dataset，其中包含 10 列。前两列是我的特征，其余是我的目标，它们可以单独放入模型中，因为它们彼此无关。因此，该问题可以作为正常的单目标回归处理。由于我需要用一个方程来描述我的输入和一些常数之间的关系，我的模型选择范围非常有限，我使用了一个最近开发的模型，称为 Equation Learner Network。

然而，我的主要问题是，尽管模型的损失非常低，例如 @Test void maxUserException() { int maxUsers = 5; MaxUserException exception = new MaxUserException(maxUsers); String message = "Excedido el número máximo de "+ maxUsers +" dispositivos"; TestUtils.testException(exception,message,ErrorCodes.MAX_USERS_DEVICES,403); } 或有时 5e-05 在 10000 个时期后，其预测非常糟糕，如此 csv file 所示.然后，我尝试通过keras ANN来解决这个问题，但结果比之前的模型差。

如果您能帮我解决错误预测的问题，我将不胜感激。

预处理步骤：

2e-05

模型类：

data = np.genfromtxt('/content/drive/MyDrive/SAEdsdata/MainData.dat',skip_header =  1)
x = data[:,:2]
y = data[:,2:]
 
imp = SimpleImputer(missing_values = np.nan,strategy = 'mean')
imp.fit(x)
x = imp.transform(x)
 
x_train,x_test,y_train,y_test = train_test_split(x,y,test_size = 0.2,random_state = 0)
 
mmsc_x = MinMaxScaler()
mmsc_z1 = MinMaxScaler()
mmsc_z2 = MinMaxScaler()
mmsc_z3 = MinMaxScaler()
mmsc_z4 = MinMaxScaler()
mmsc_z5 = MinMaxScaler()
mmsc_pred = MinMaxScaler()
 
x_trainfit = mmsc_x.fit_transform(x_train)
x_testfit = mmsc_x.transform(x_test)
 
#y_train = np.array(y_train,ndmin = 2)
#y_test = np.array(y_test,ndmin = 2)
 
 
#z in the third column
z1_train = y_train[:,0]
z1_train = z1_train.reshape(len(z1_train),1)
z1_trainfit = mmsc_z1.fit_transform(z1_train)
z1_test = y_test[:,0]
z1_test = z1_test.reshape(len(z1_test),1)
z1_testfit = mmsc_z1.transform(z1_test)
 
#z in the fourth column
z2_train = y_train[:,1]
z2_train = z2_train.reshape(len(z2_train),1)
z2_trainfit = mmsc_z2.fit_transform(z2_train)
z2_test = y_test[:,1]
z2_test = z2_test.reshape(len(z2_test),1)
z2_testfit = mmsc_z2.transform(z2_test)
 
#z in the eighth column 
z3_train = y_train[:,5]
z3_train = z3_train.reshape(len(z3_train),1)
z3_trainfit = mmsc_z3.fit_transform(z3_train)
z3_test = y_test[:,5]
z3_test = z3_test.reshape(len(z3_test),1)
z3_testfit = mmsc_z3.transform(z3_test)
 
#z in the nineth column 
z4_train = y_train[:,6]
z4_train = z4_train.reshape(len(z4_train),1)
z4_trainfit = mmsc_z4.fit_transform(z4_train)
z4_test = y_test[:,6]
z4_test = z4_test.reshape(len(z4_test),1)
z4_testfit = mmsc_z4.transform(z4_test)
 
#z in the tenth column 
z5_train = y_train[:,7]
z5_train = z5_train.reshape(len(z5_train),1)
z5_trainfit = mmsc_z5.fit_transform(z5_train)
z5_test = y_test[:,7]
z5_test = z5_test.reshape(len(z5_test),1)
z5_testfit = mmsc_z5.transform(z5_test)

方法调用：

class EQLModel:
  def __init__(self):
    self.eqlModel = EQL()
    self.z_pred = []
    self.z_pred_inversed = []
 
  def model_configuration_and_training(self,z,hidden_layers = 4,epoch = 5000,batch = 32,step_per_epoch = None):
    self.eqlModel = EQL(hidden_layers,dim = 2)
    self.eqlModel.build_and_compile_model()
    self.eqlModel.fit(x_trainfit,lmbda = ['x','y'],t0 = epoch,batch_size = batch,shuffle = False,verbose = 1,steps_per_epoch = step_per_epoch)
      
  def single_prediction(self,prediction_x):
    predicted = self.eqlModel.predict(mmsc_x.transform(prediction_x)) 
    predicted_value = mmsc_z1.inverse_transform(predicted.reshape(1,-1)).ravel()
    print("{:.6e}".format(predicted_value[0]))
 
  def test_set_prediction(self,z_test,z_scaler_class,z_testfit):
    self.z_pred = self.eqlModel.predict(x_testfit)
    self.z_pred_inversed = z_scaler_class.inverse_transform(self.z_pred.reshape(-1,1))
    print('R2 score: {:.2f}'.format(r2_score(z_test,self.z_pred_inversed)))
    mse = mean_squared_error(z_testfit,self.z_pred)
    print('RMSE: {:.2f}'.format(mse))
  
  def make_csv(self,name,z_testfit):
    with open(name + '.csv','w',newline = '') as f:
      thewriter = csv.writer(f)
      thewriter.writerow(['x','y','z','actual z with inversing','our inversed predictions','z without inversing','our predictions without invesing','error percentage'])
      for i in range(len(x_test)):    
        ws_x = '{:.6e}'.format(x_test[i][0])
        ws_y = '{:.6e}'.format(x_test[i][1])
        ws_z = '{:.6e}'.format(z_test[i][0])
        z_with_i_array = z_scaler_class.inverse_transform(z_testfit)
        z_with_i = '{:.6e}'.format(z_with_i_array[i][0])
        zp_with_i = '{:.6e}'.format(self.z_pred_inversed[i][0])    
        z_without_i = '{:.6e}'.format(z_testfit[i][0])
        zp_without_i = '{:.6e}'.format(self.z_pred[i][0])
        error = '{:.2f}'.format(np.abs(np.abs(z_test[i][0] - self.z_pred_inversed[i][0]) / z_test[i][0] * 100))
        thewriter.writerow([ws_x,ws_y,ws_z,z_with_i,zp_with_i,z_without_i,zp_without_i,error])
 
  def evaluate_model(self,plot = True,batch = 32):
    history = self.eqlModel.evaluate(x_trainfit,verbose = 1)
    print(history)
    #if plot:
 
eqlm = EQLModel()

人工神经网络配置：

eqlm.model_configuration_and_training(z = z1_trainfit,hidden_layers = 5,batch = 512,epoch = 10000)
eqlm.single_prediction([[1.357238e+00,1.614171e+04]])
eqlm.test_set_prediction(z1_test,mmsc_z1,z1_testfit)
eqlm.make_csv(name = 'z1perf',z_test = z1_test,z_scaler_class = mmsc_z1,z_testfit = z1_testfit)

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）

keraskerasneural-networknon-linear-regressionpython