0

tl; dr 我尝试使用 Hyperas 优化和交叉验证我的超参数，但无法使用 KerasClassifier 进行预处理（缩放、过采样/欠采样）管道

我使用Hyperas（hyperopt 的包装器）来调整我的神经网络（使用 Keras/Tensorflow 构建）的超参数，并尝试实施 kfold-cross 验证以获得最佳参数。但是，我还对数据进行预处理（Standardscaler 和 MinMaxScaler），然后使用 SMOTETOMEK 进行过采样/欠采样）。

我读到不应该对整个数据集进行特征缩放和重采样，而应该只对用于训练的部分进行特征缩放和重采样，以避免溢出。尝试在 hyperopt 中仅针对交叉验证的 train folds 实现这一点有些困难，因为当使用类似 的管道时imblearn，该管道仅适用于仅采用模型函数的 KerasClassifier。我不能给他那个模型函数，因为 hyperopt 中的整个验证过程都发生在一个函数中。

你对如何制作这样的作品有什么建议吗？我可以def data()在整个数据集上进行所有预处理并优化/交叉验证参数吗？或者这会损害正确的参数查找过程吗？（我确实有最终模型的额外测试数据集）

有没有办法让它手动工作？

def data():
    import pandas as pd
    import feather

    df_hyper_X = feather.read_dataframe('df_hyper_X_train.feather')
    df_hyper_Y = feather.read_dataframe('df_hyper_Y_train.feather')

    return df_hyper_X, df_hyper_Y

def hyper_model(df_hyper_X,df_hyper_Y):

  stdscl_features = ['pre_grade', 'math']
  normscl_features = 'time'
  stdscl_transformer = Pipeline(steps=[('stdscaler', StandardScaler())])
  normscl_transformer = Pipeline(steps=[('normscaler', MinMaxScaler())])

  preprocessor = ColumnTransformer(transformers=[('stdscl', stdscl_transformer, stdscl_features),('minmaxscl', normscl_transformer, normscl_features)], remainder='passthrough')

  metrics = [
            tf.keras.metrics.TruePositives(name='tp'),
            tf.keras.metrics.FalsePositives(name='fp'),
            tf.keras.metrics.TrueNegatives(name='tn'),
            tf.keras.metrics.FalseNegatives(name='fn'), 
            tf.keras.metrics.BinaryAccuracy(name='accuracy'),
            tf.keras.metrics.Precision(name='precision'),
            tf.keras.metrics.AUC(name='auc'),
             ]

  model = tf.keras.Sequential()
  model.add(Dense({{choice([2,4,8,16,32,64])}}, activation={{choice(['relu', 'sigmoid', 'tanh', 'elu', 'selu'])}}, kernel_initializer={{choice(['lecun_uniform','glorot_normal', 'glorot_uniform', 'he_normal', 'he_uniform'])}}
                  , input_shape=(16,))) #If ReLu use --> HE uniform initialization #kernel_regularizer=tf.keras.regularizers.l2({{choice([0.01, 0.05, 0.1])}}
          #model.add(LeakyReLU(alpha={{uniform(0.5, 1)}}))
  model.add(Dropout({{uniform(0, 1)}}))      
  if ({{choice(['one', 'two'])}}) == 'two':
      model.add(Dense({{choice([2,4,8,16,32,64])}}, activation={{choice(['relu', 'sigmoid', 'tanh', 'elu', 'selu'])}}))
      model.add(Dropout({{uniform(0, 1)}}))

  #model.add(Dense({{choice([2,4,8,16,32,64])}}, activation={{choice(['relu', 'sigmoid', 'tanh', 'elu', 'selu'])}})) third hidden layer
  #model.add(Dropout({{uniform(0, 1)}}))

  model.add(Dense(1, activation='sigmoid'))

  adam = tf.keras.optimizers.Adam(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  nadam = tf.keras.optimizers.Nadam(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  adamax = tf.keras.optimizers.Adamax(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  adagrad = tf.keras.optimizers.Adagrad(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  adadelta = tf.keras.optimizers.Adadelta(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  sgd = tf.keras.optimizers.SGD(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  rmsprop = tf.keras.optimizers.RMSprop(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})


  opti_choice = {{choice(['adam', 'nadam', 'adamax','adagrad', 'adadelta', 'sgd','rmsprop'])}}
  if opti_choice == 'adam':
      optimizer = adam
  elif opti_choice == 'nadam':
      optimizer = nadam
  elif opti_choice == 'adamax':
      optimizer = adamax
  elif opti_choice == 'adagrad':
      optimizer = adagrad
  elif opti_choice == 'adadelta':
      optimizer = adadelta
  elif opti_choice == 'sgd':
      optimizer = sgd
  else:
      optimizer = rmsprop

  model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=metrics)

  kfold = KFold(n_splits=10, shuffle=True, random_state=3)

  imba_pipeline = make_pipeline(preprocessor, SMOTETomek(sampling_strategy='auto', random_state=2),
                                KerasClassifier(model, epochs={{choice([20,30,40,50,60,70])}}, batch_size={{choice([16,32, 64, 128])}}, verbose=0))
  results = cross_val_score(imba_pipeline, df_hyper_X, df_hyper_Y, cv=kfold, scoring='precision').mean()


  print('Precision', results)
  return {'loss': -results, 'status': STATUS_OK, 'model': model}

if __name__ == '__main__':
    best_run, best_model = optim.minimize(model=hyper_model,
                                          data=data,
                                          algo=tpe.suggest,
                                          max_evals=30,
                                          trials=Trials(),
                                          notebook_name = 'drive/My Drive/Colab Notebooks/final_NL_EU_Non-EU')
    X_train, Y_train, X_test, Y_test = data()
    print("Evalutation of best performing model:")
    print(best_model.evaluate(X_test, Y_test))
    print("Best performing model chosen hyper-parameters:")
    print(best_run)
4

1 回答 1

1

解决了。如果有人感兴趣，这是解决方案：

def data():
    import pandas as pd
    import feather

    df_hyper_X = feather.read_dataframe('df_hyper_X_train.feather')
    df_hyper_Y = feather.read_dataframe('df_hyper_Y_train.feather')

    return df_hyper_X, df_hyper_Y

def hyper_model(df_hyper_X,df_hyper_Y):

  ct = ColumnTransformer([('ct_std', StandardScaler(), ['pre_grade', 'math']),('ct_minmax', MinMaxScaler(), ['time'])
  ], remainder='passthrough')

  metrics = [
            tf.keras.metrics.TruePositives(name='tp'),
            tf.keras.metrics.FalsePositives(name='fp'),
            tf.keras.metrics.TrueNegatives(name='tn'),
            tf.keras.metrics.FalseNegatives(name='fn'), 
            tf.keras.metrics.BinaryAccuracy(name='accuracy'),
            tf.keras.metrics.Precision(name='precision'),
            tf.keras.metrics.AUC(name='auc'),
             ]

  model = tf.keras.Sequential()
  model.add(Dense({{choice([2,4,8,16,32,64])}}, activation={{choice(['relu', 'sigmoid', 'tanh', 'elu', 'selu'])}}, kernel_initializer={{choice(['lecun_uniform','glorot_normal', 'glorot_uniform', 'he_normal', 'he_uniform'])}}
                  , input_shape=(20,)))
  model.add(Dropout({{uniform(0, 0.5)}}))

  if ({{choice(['one', 'two'])}}) == 'two':
      model.add(Dense({{choice([2,4,8,16,32,64])}}, activation={{choice(['relu', 'sigmoid', 'tanh', 'elu', 'selu'])}}))
      model.add(Dropout({{uniform(0, 0.5)}}))

  model.add(Dense(1, activation='sigmoid'))

  adam = tf.keras.optimizers.Adam(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  nadam = tf.keras.optimizers.Nadam(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  adamax = tf.keras.optimizers.Adamax(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  adagrad = tf.keras.optimizers.Adagrad(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  adadelta = tf.keras.optimizers.Adadelta(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  sgd = tf.keras.optimizers.SGD(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})
  rmsprop = tf.keras.optimizers.RMSprop(lr={{choice([0.0001, 0.001, 0.01, 0.1])}})

  opti_choice = {{choice(['adam', 'nadam', 'adamax','adagrad', 'adadelta', 'sgd','rmsprop'])}}
  if opti_choice == 'adam':
      optimizer = adam
  elif opti_choice == 'nadam':
      optimizer = nadam
  elif opti_choice == 'adamax':
      optimizer = adamax
  elif opti_choice == 'adagrad':
      optimizer = adagrad
  elif opti_choice == 'adadelta':
      optimizer = adadelta
  elif opti_choice == 'sgd':
      optimizer = sgd
  else:
      optimizer = rmsprop

  model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=metrics)

  smt = SMOTETomek(sampling_strategy='auto', random_state=2)
  kfold = KFold(n_splits=10, shuffle=True, random_state=3)  
  scores = []

  for train_fold_index, val_fold_index in kfold.split(df_hyper_X,df_hyper_Y):

    X_train_fold, y_train_fold = df_hyper_X.iloc[train_fold_index], df_hyper_Y.iloc[train_fold_index]

    X_val_fold, y_val_fold = df_hyper_X.iloc[val_fold_index], df_hyper_Y.iloc[val_fold_index]

    X_train_fold = ct.fit_transform(X_train_fold)
    X_val_fold = ct.transform(X_val_fold)

    X_train_smtk, y_train_smtk = smt.fit_resample(X_train_fold, y_train_fold)

    model.fit(X_train_smtk, y_train_smtk, epochs={{choice([20,30,40,50,60,70])}}, batch_size={{choice([16,32, 64, 128])}})

    predicts = model.predict(X_val_fold)
    score = precision_score(y_val_fold, predicts.round())
    scores.append(score)

  avg_score = np.mean(scores)    
  print('Precision', avg_score)
  return {'loss': -avg_score, 'status': STATUS_OK, 'model': model}

if __name__ == '__main__':
    best_run, best_model = optim.minimize(model=hyper_model,
                                          data=data,
                                          algo=tpe.suggest,
                                          max_evals=2,
                                          trials=Trials(),
                                          notebook_name = 'drive/My Drive/Colab Notebooks/final_NL_EU_Non-EU')
    df_hyper_X, df_hyper_Y = data()
    print("Best performing model chosen hyper-parameters:")
    print(best_run)
于 2020-05-19T21:25:29.697 回答