CODEWITHSUNDEEP
pythontensorflowkerasscikit-learnneural-network
Julieta Cadete
Julieta Cadete

Reputation: 21

ValueError: activation is not a legal parameter

I'm trying to tune my model but I'm getting this Value Error. I tried to change the activation function, but when I did the learning rate returned the same error. I'm not sure if I'm missing something.

>ValueError                              Traceback (most recent call last)

> <ipython-input-46-5d07e2ad456a> in <module>
>      9   param_distributions = params,
>     10   cv = KFold(10))

> --->11 random_search_results = random_search.fit(X_train, y_train)

ValueError: activation is not a legal parameter

def create_model(learning_rate=0.01):
    opt = 'Adam'
    Tuning_model = Sequential()
    Tuning_model.add(Dense(16, input_shape=(X_train.shape[1],)))
    Tuning_model.add(Dropout(.2))
    Tuning_model.add(BatchNormalization())
    Tuning_model.add(Activation('relu'))
    Tuning_model.add(Dense(32))
    Tuning_model.add(Dropout(.2))
    Tuning_model.add(Dense(1))
    Tuning_model.compile(loss='mse', optimizer=opt, metrics='mse')
    
    return Tuning_model

# Define the hyperparameter space
params = {'activation': ["relu", "tanh"],
          'batch_size': [16, 32, 64, 128], 
          'epochs': [50, 100],
          'optimizer': ["Adam", "SGD", "RMSprop"],
          'learning_rate': [0.01, 0.001, 0.0001]}
# Create a randomize search cv object 
random_search = RandomizedSearchCV(Tuning_model,
                                   param_distributions = params,
                                   cv = KFold(10))
random_search_results = random_search.fit(X_train, y_train)

Upvotes: 2

Views: 1002

Answers (1)

Poe Dator
Poe Dator

Reputation: 4893

The ValueError is raised because activation is not a parameter of the whole model, but rather a parameter of certain of its' layer(s). So when RandomizedSearchCV tries to pass it, Model object can't accept it.

I suggest 2 solutions:

  1. use KerasClassifier wrapper around build function and make activation one of its. Then optimize its performance with RandomizedSearchCV.
  1. Use a dedicated optimization package, like optuna - it works smarter and has greater degree of customization. Try it - they have good and easy docs on their site.

Side note: RandomizedSearchCV with 10 folds is an overkill, if sample is large enough, make it 2 or even single fold.

Upvotes: 4

Related Questions