I'm a beginner in machine learning and want to train a CNN (for image recognition) with optimized hyperparameter like dropout rate, learning rate and number of epochs.

The optimal hyperparameter I try to find via GridSearchCV from Scikit-learn. I have often read that GridSearchCV can be used in combination with early stopping, but I can not find a sample code in which this is demonstrated.

With EarlyStopping I would try to find the optimal number of epochs, but I don't know how I can combine EarlyStopping with GridSearchCV or at least with cross validation.

Can anyone give me a hint on how to do that, it would be a great help?

My current code looks like this:

def create_model(dropout_rate_1=0.0, dropout_rate_2=0.0, learn_rate=0.001):
    model = Sequential()
    model.add(Conv2D(32, kernel_size=(3,3), input_shape=(28,28,1), activation='relu', padding='same')
    model.add(Conv2D(32, kernel_size=(3,3), activation='relu', padding='same')
    model.add(Dense(128, activation='relu'))
    model.add(Dense(10, activation='softmax'))
    model.compile(loss='categorical_crossentropy', optimizer=optimizer, 
    return model

model = KerasClassifier(build_fn=create_model, epochs=50, batch_size=10, verbose=0)
epochs = [30, 40, 50, 60]
dropout_rate_1 = [0.0, 0.2, 0.4, 0.6]
dropout_rate_2 = [0.0, 0.2, 0.4, 0.6]
learn_rate = [0.0001, 0.001, 0.01]
param_grid = dict(dropout_rate_1=dropout_rate_1, dropout_rate_2=dropout_rate_2,
                        learn_rate=learn_rate, epochs=epochs)
grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=5)
grid_result = grid.fit(X, y) 

  • $\begingroup$ Can you add some code? What exactly are you trying to achieve. Early stopping usually means that if, after x steps, no progress is achieved, you try a different set of parameters. So it usually means to set a cap on the number of attempts to optimize with a given parameter set. $\endgroup$
    – Peter
    Nov 15, 2019 at 22:13
  • $\begingroup$ @Peter sorry, I've just discovered your answer. Current code has been inserted above. $\endgroup$
    – Code Now
    Nov 16, 2019 at 11:27

3 Answers 3


Just to add to others here. I guess you simply need to include a early stopping callback in your fit().

Something like:

from keras.callbacks import EarlyStopping
# Define early stopping
early_stopping = EarlyStopping(monitor='val_loss', patience=epochs_to_wait_for_improve)
# Add ES into fit
history = model.fit(..., callbacks=[early_stopping])

If you would ask for code suggestion please specify your framework in the future. I am assuming you are using Keras

I can make you a minimum viable implementation of your case.

from sklearn.base import ClassifierMixin, BaseEstimator

class CNN_model(ClassifierMixin, BaseEstimator) :
      def __init__(**model_params) :
              define model parameters within this init function
          self.model = # Use the params above and make a keras model and store it in this variable
          self.model.compile(loss=  , optimizer =   , metrics=[]) # Please fill-in the appropriate loss and metrics

       def fit(X,y) :
          self.model.fit(X,y, training_params)
          # You specify everything in training_params e.g. epoch, callbacks(which includes early stopping)
          return self

       def predict(X) :
          return self.model.predict(X)

So basically the code above makes a custom instance of sklearn estimator, which if you are succesfully build can be combined with GridSearchCV.


GridSearchCv with Early Stopping - I was curious about your question. As long as the algorithms has built in Early Stopper feature, you can use it in this manner.

when it comes to other algorithms, It might not serve the purpose of early stopping because you never know what parameters are gonna be the best until you experiment with them.

  • $\begingroup$ The idea would be to replace the search area "epochs" with early stopping so that each fold stops as soon as validation_loss increases (considering patience). My goal would be to shorten the time-consuming search by GridSearch. $\endgroup$
    – Code Now
    Nov 16, 2019 at 11:42
  • $\begingroup$ I am with you on that Early stopping would be awesome. But it is what it is. the Randomness, entropy, information gain and uncertainity etc.. of data, it becomes difficult to where the data would converge, descent, ascend or diverge. End of the data it works but with more time and computation :) $\endgroup$
    – Syenix
    Nov 25, 2019 at 10:39
  • $\begingroup$ I wonder what the normal way of handling with Early stopping would look like. Suppose I have determined the optimal hyperparameters via GridSearchCV (without using of Early Stopping and without searching for optimal number of epochs). Would Early stopping be recommended for the training of the CNN with the optimized hyperparameters? If so, would not I have to reserve a validation set for EarlyStopping before running GridSearchCV? $\endgroup$
    – Code Now
    Nov 26, 2019 at 1:47
  • $\begingroup$ I see, Early stopping is available in Tensorflow and Pytorch if you want to train the CNN. For each epoch, the loss is calculated and once the loss is saturated. the execution stops. You dont have to worry when you switch to CNN using Keras and Tensorflow or Pytorch. :) $\endgroup$
    – Syenix
    Nov 26, 2019 at 5:13

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.