How to Grid Search Hyperparameters for Deep Learning Models in Python With Keras

Hyperparameter optimization is a big part of deep learning.

The reason is that neural networks are notoriously difficult to configure and there are a lot of parameters that need to be set. On top of that, individual models can be very slow to train.

In this post you will discover how you can use the grid search capability from the scikit-learn python machine learning library to tune the hyperparameters of Keras deep learning models.

After reading this post you will know:

  • How to wrap Keras models for use in scikit-learn and how to use grid search.
  • How to grid search common neural network parameters such as learning rate, dropout rate, epochs and number of neurons.
  • How to define your own hyperparameter tuning experiments on your own projects.

Let’s get started.

  • Update Nov/2016: Fixed minor issue in displaying grid search results in code examples.
  • Update Oct/2016: Updated examples for Keras 1.1.0, TensorFlow 0.10.0 and scikit-learn v0.18.
  • Update Mar/2017: Updated example for Keras 2.0.2, TensorFlow 1.0.1 and Theano 0.9.0.
  • Update Sept/2017: Updated example to use Keras 2 “epochs” instead of Keras 1 “nb_epochs”.
How to Grid Search Hyperparameters for Deep Learning Models in Python With Keras

How to Grid Search Hyperparameters for Deep Learning Models in Python With Keras
Photo by 3V Photo, some rights reserved.


In this post, I want to show you both how you can use the scikit-learn grid search capability and give you a suite of examples that you can copy-and-paste into your own project as a starting point.

Below is a list of the topics we are going to cover:

  1. How to use Keras models in scikit-learn.
  2. How to use grid search in scikit-learn.
  3. How to tune batch size and training epochs.
  4. How to tune optimization algorithms.
  5. How to tune learning rate and momentum.
  6. How to tune network weight initialization.
  7. How to tune activation functions.
  8. How to tune dropout regularization.
  9. How to tune the number of neurons in the hidden layer.

How to Use Keras Models in scikit-learn

Keras models can be used in scikit-learn by wrapping them with the KerasClassifier or KerasRegressor class.

To use these wrappers you must define a function that creates and returns your Keras sequential model, then pass this function to the build_fn argument when constructing the KerasClassifier class.

For example:

The constructor for the KerasClassifier class can take default arguments that are passed on to the calls to, such as the number of epochs and the batch size.

For example:

The constructor for the KerasClassifier class can also take new arguments that can be passed to your custom create_model() function. These new arguments must also be defined in the signature of your create_model() function with default parameters.

For example:

You can learn more about the scikit-learn wrapper in Keras API documentation.

How to Use Grid Search in scikit-learn

Grid search is a model hyperparameter optimization technique.

In scikit-learn this technique is provided in the GridSearchCV class.

When constructing this class you must provide a dictionary of hyperparameters to evaluate in the param_grid argument. This is a map of the model parameter name and an array of values to try.

By default, accuracy is the score that is optimized, but other scores can be specified in the score argument of the GridSearchCV constructor.

By default, the grid search will only use one thread. By setting the n_jobs argument in the GridSearchCV constructor to -1, the process will use all cores on your machine. Depending on your Keras backend, this may interfere with the main neural network training process.

The GridSearchCV process will then construct and evaluate one model for each combination of parameters. Cross validation is used to evaluate each individual model and the default of 3-fold cross validation is used, although this can be overridden by specifying the cv argument to the GridSearchCV constructor.

Below is an example of defining a simple grid search:

Once completed, you can access the outcome of the grid search in the result object returned from The best_score_ member provides access to the best score observed during the optimization procedure and the best_params_ describes the combination of parameters that achieved the best results.

You can learn more about the GridSearchCV class in the scikit-learn API documentation.

Problem Description

Now that we know how to use Keras models with scikit-learn and how to use grid search in scikit-learn, let’s look at a bunch of examples.

All examples will be demonstrated on a small standard machine learning dataset called the Pima Indians onset of diabetes classification dataset. This is a small dataset with all numerical attributes that is easy to work with.

  1. Download the dataset and place it in your currently working directly with the name pima-indians-diabetes.csv.

As we proceed through the examples in this post, we will aggregate the best parameters. This is not the best way to grid search because parameters can interact, but it is good for demonstration purposes.

Note on Parallelizing Grid Search

All examples are configured to use parallelism (n_jobs=-1).

If you get an error like the one below:

Kill the process and change the code to not perform the grid search in parallel, set n_jobs=1.

Need help with Deep Learning in Python?

Take my free 2-week email course and discover MLPs, CNNs and LSTMs (with sample code).

Click to sign-up now and also get a free PDF Ebook version of the course.

Start Your FREE Mini-Course Now!

How to Tune Batch Size and Number of Epochs

In this first simple example, we look at tuning the batch size and number of epochs used when fitting the network.

The batch size in iterative gradient descent is the number of patterns shown to the network before the weights are updated. It is also an optimization in the training of the network, defining how many patterns to read at a time and keep in memory.

The number of epochs is the number of times that the entire training dataset is shown to the network during training. Some networks are sensitive to the batch size, such as LSTM recurrent neural networks and Convolutional Neural Networks.

Here we will evaluate a suite of different mini batch sizes from 10 to 100 in steps of 20.

The full code listing is provided below.

Running this example produces the following output.

We can see that the batch size of 20 and 100 epochs achieved the best result of about 68% accuracy.

How to Tune the Training Optimization Algorithm

Keras offers a suite of different state-of-the-art optimization algorithms.

In this example, we tune the optimization algorithm used to train the network, each with default parameters.

This is an odd example, because often you will choose one approach a priori and instead focus on tuning its parameters on your problem (e.g. see the next example).

Here we will evaluate the suite of optimization algorithms supported by the Keras API.

The full code listing is provided below.

Running this example produces the following output.

The results suggest that the ADAM optimization algorithm is the best with a score of about 70% accuracy.

How to Tune Learning Rate and Momentum

It is common to pre-select an optimization algorithm to train your network and tune its parameters.

By far the most common optimization algorithm is plain old Stochastic Gradient Descent (SGD) because it is so well understood. In this example, we will look at optimizing the SGD learning rate and momentum parameters.

Learning rate controls how much to update the weight at the end of each batch and the momentum controls how much to let the previous update influence the current weight update.

We will try a suite of small standard learning rates and a momentum values from 0.2 to 0.8 in steps of 0.2, as well as 0.9 (because it can be a popular value in practice).

Generally, it is a good idea to also include the number of epochs in an optimization like this as there is a dependency between the amount of learning per batch (learning rate), the number of updates per epoch (batch size) and the number of epochs.

The full code listing is provided below.

Running this example produces the following output.

We can see that relatively SGD is not very good on this problem, nevertheless best results were achieved using a learning rate of 0.01 and a momentum of 0.0 with an accuracy of about 68%.

How to Tune Network Weight Initialization

Neural network weight initialization used to be simple: use small random values.

Now there is a suite of different techniques to choose from. Keras provides a laundry list.

In this example, we will look at tuning the selection of network weight initialization by evaluating all of the available techniques.

We will use the same weight initialization method on each layer. Ideally, it may be better to use different weight initialization schemes according to the activation function used on each layer. In the example below we use rectifier for the hidden layer. We use sigmoid for the output layer because the predictions are binary.

The full code listing is provided below.

Running this example produces the following output.

We can see that the best results were achieved with a uniform weight initialization scheme achieving a performance of about 72%.

How to Tune the Neuron Activation Function

The activation function controls the non-linearity of individual neurons and when to fire.

Generally, the rectifier activation function is the most popular, but it used to be the sigmoid and the tanh functions and these functions may still be more suitable for different problems.

In this example, we will evaluate the suite of different activation functions available in Keras. We will only use these functions in the hidden layer, as we require a sigmoid activation function in the output for the binary classification problem.

Generally, it is a good idea to prepare data to the range of the different transfer functions, which we will not do in this case.

The full code listing is provided below.

Running this example produces the following output.

Surprisingly (to me at least), the ‘linear’ activation function achieved the best results with an accuracy of about 72%.

How to Tune Dropout Regularization

In this example, we will look at tuning the dropout rate for regularization in an effort to limit overfitting and improve the model’s ability to generalize.

To get good results, dropout is best combined with a weight constraint such as the max norm constraint.

For more on using dropout in deep learning models with Keras see the post:

This involves fitting both the dropout percentage and the weight constraint. We will try dropout percentages between 0.0 and 0.9 (1.0 does not make sense) and maxnorm weight constraint values between 0 and 5.

The full code listing is provided below.

Running this example produces the following output.

We can see that the dropout rate of 0.2% and the maxnorm weight constraint of 4 resulted in the best accuracy of about 72%.

How to Tune the Number of Neurons in the Hidden Layer

The number of neurons in a layer is an important parameter to tune. Generally the number of neurons in a layer controls the representational capacity of the network, at least at that point in the topology.

Also, generally, a large enough single layer network can approximate any other neural network, at least in theory.

In this example, we will look at tuning the number of neurons in a single hidden layer. We will try values from 1 to 30 in steps of 5.

A larger network requires more training and at least the batch size and number of epochs should ideally be optimized with the number of neurons.

The full code listing is provided below.

Running this example produces the following output.

We can see that the best results were achieved with a network with 5 neurons in the hidden layer with an accuracy of about 71%.

Tips for Hyperparameter Optimization

This section lists some handy tips to consider when tuning hyperparameters of your neural network.

  • k-fold Cross Validation. You can see that the results from the examples in this post show some variance. A default cross-validation of 3 was used, but perhaps k=5 or k=10 would be more stable. Carefully choose your cross validation configuration to ensure your results are stable.
  • Review the Whole Grid. Do not just focus on the best result, review the whole grid of results and look for trends to support configuration decisions.
  • Parallelize. Use all your cores if you can, neural networks are slow to train and we often want to try a lot of different parameters. Consider spinning up a lot of AWS instances.
  • Use a Sample of Your Dataset. Because networks are slow to train, try training them on a smaller sample of your training dataset, just to get an idea of general directions of parameters rather than optimal configurations.
  • Start with Coarse Grids. Start with coarse-grained grids and zoom into finer grained grids once you can narrow the scope.
  • Do not Transfer Results. Results are generally problem specific. Try to avoid favorite configurations on each new problem that you see. It is unlikely that optimal results you discover on one problem will transfer to your next project. Instead look for broader trends like number of layers or relationships between parameters.
  • Reproducibility is a Problem. Although we set the seed for the random number generator in NumPy, the results are not 100% reproducible. There is more to reproducibility when grid searching wrapped Keras models than is presented in this post.


In this post, you discovered how you can tune the hyperparameters of your deep learning networks in Python using Keras and scikit-learn.

Specifically, you learned:

  • How to wrap Keras models for use in scikit-learn and how to use grid search.
  • How to grid search a suite of different standard neural network parameters for Keras models.
  • How to design your own hyperparameter optimization experiments.

Do you have any experience tuning hyperparameters of large neural networks? Please share your stories below.

Do you have any questions about hyperparameter optimization of neural networks or about this post? Ask your questions in the comments and I will do my best to answer.

Frustrated With Your Progress In Deep Learning?

Deep Learning with Python

 What If You Could Develop A Network in Minutes

…with just a few lines of Python

Discover how in my new Ebook: Deep Learning With Python

It covers self-study tutorials and end-to-end projects on topics like:
Multilayer PerceptronsConvolutional Nets and Recurrent Neural Nets, and more…

Finally Bring Deep Learning To
Your Own Projects

Skip the Academics. Just Results.

Click to learn more.

180 Responses to How to Grid Search Hyperparameters for Deep Learning Models in Python With Keras

  1. Yanbo August 9, 2016 at 9:10 am #

    As always excellent post,. I’ve been doing some hyper-parameter optimization by hand, but I’ll definitely give Grid Search a try.

    Is it possible to set up a different threshold for sigmoid output in Keras? Rather then using 0.5 I was thinking of trying 0.7 or 0.8

    • Jason Brownlee August 15, 2016 at 11:10 am #

      Thanks Yanbo.

      I don’t think so, but you could implement your own activation function and do anything you wish.

      • Shudhan September 5, 2016 at 6:20 pm #

        My question is related to this thread. How to get the probablities as the output? I dont want the class output. I read for a regression problem that no activation function is needed in the output layer. Similiar implementation will get me the probabilities ?? or the output will exceed 0 and 1??

        • Jason Brownlee September 6, 2016 at 9:41 am #

          Hi Shudhan, you can use a sigmoid activation and treat the outputs like probabilities (they will be in the range of 0-1).

  2. eclipsedu August 18, 2016 at 5:55 pm #

    Sound awesome!Will this grid search method use the full cpu(which can be 8/16 cores) ?

  3. Reza August 18, 2016 at 6:00 pm #

    Great post,
    Can I use this tips on CNNs in keras as well?

    • Jason Brownlee August 19, 2016 at 5:24 am #

      They can be a start, but remember it is a good idea to use a repeating structure in a large CNN and you will need to tune the number of filters and pool size.

  4. Prashant August 22, 2016 at 4:55 pm #

    Hi Jason, First of all great post! I applied this by dividing the data into train and test and used train dataset for grid fit. Plan was to capture best parameters in train and apply them on test to see accuracy. But it seems and applied with same parameters on same dataset (in this case train) give different accuracy results. Any idea why this happens. I can share the code if it helps.

    • Jason Brownlee August 23, 2016 at 6:00 am #

      You will see small variation in the performance of a neural net with the same parameters from run to run. This is because of the stochastic nature of the technique and how very hard it is to fix the random number seed successfully in python/numpy/theano.

      You will also see small variation due to the data used to train the method.

      Generally, you could use all of your data to grid search to try to reduce the second type of variation (slower). You could store results and use statistical significance tests to compare populations of results to see if differences are significant to sort out the first type or variation.

      I hope that helps.

  5. vinay August 22, 2016 at 9:05 pm #

    hi, I think this will best tutorial i ever found on web….Thanks for sharing….is it possible to use these tips on LSTM, Bilstm cnnlstm

    • Jason Brownlee August 23, 2016 at 5:57 am #

      Thanks Vinay, I’m glad it’s useful.

      Absolutely, you could use these tactics on other algorithm types.

  6. shudhan September 2, 2016 at 3:26 pm #

    Best place to learn the tuning.. my question – is it good to follow the order you mentioned to tune the parameters? I know the most significant parameters should be tuned first

    • Jason Brownlee September 3, 2016 at 6:56 am #

      Thanks. The order is a good start. It is best to focus on areas where you think you will get the biggest improvement first – which is often the structure of the network (layers and neurons).

  7. Satheesh September 27, 2016 at 12:24 am #

    when I am using the categorical_entropy loss function and running the grid search with n_jobs more than 1 its throwing error “cannot pickle object class”, but the same thing is working fine with binary_entropyloss. Can you tell me if I am making any mistake in my code:
    def create_model(optimizer=’adam’):
    # create model
    model.add(Dense(30, input_dim=59, init=’normal’, activation=’relu’))
    model.add(Dense(15, init=’normal’, activation=’sigmoid’))
    model.add(Dense(3, init=’normal’, activation=’sigmoid’))
    # Compile model
    model.compile(loss=’categorical_crossentropy’, optimizer=optimizer, metrics=[‘accuracy’])
    return model

    # Create Keras Classifier
    print “——————— Running Grid Search on Keras Classifier for epochs and batch ——————”
    clf = model = KerasClassifier(build_fn = create_model, verbose=0)
    param_grid = {“batch_size”:range(10, 30, 10), “nb_epoch”:range(50, 150, 50)}
    optimizer = [‘SGD’, ‘RMSprop’, ‘Adagrad’, ‘Adadelta’, ‘Adam’, ‘Adamax’, ‘Nadam’]
    grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=4)
    grid_result =, y_train)
    print(“Best: %f using %s” % (grid_result.best_score_, grid_result.best_params_))

    • Jason Brownlee September 27, 2016 at 7:44 am #

      Strange Satheesh, I have not seen that before.

      Let me know if you figure it out.

      • Kai September 18, 2017 at 10:01 pm #

        I came cross and solved the problem several days ago. Please use “epochs” instead of “nb_epoch” in param_grid dict. Personally, I guess “cannot pickle object class” means the neuron network cannot be built because of some errors. Open to discussion.

        • Jason Brownlee September 19, 2017 at 7:40 am #

          Glad to hear it.

          I updated the example to use “epochs” to work with Keras 2.

  8. L Fenu November 9, 2016 at 7:47 pm #

    excellent post, thanks. It’s been very helpful to get me started on hyperparameterisation.

    One thing I haven’t been able to do yet is to grid search over parameters which are not proper to the NN but to the trainign set. For example, I can fine-tune the input_dim parameter by creating a function generator which takes care of creating the function that will create the model, like this:

    # fp_subset is a subset of columns of my whole training set.

    create_basic_ANN_model = kt.ANN_model_gen( # defined elsewhere
    input_dim=len(fp_subset), output_dim=1, layers_num=2, layers_sizes=[len(fp_subset)/5, len(fp_subset)/10, ],
    loss=’mean_squared_error’, optimizer=’adadelta’, metrics=[‘mean_squared_error’, ‘mean_absolute_error’]

    model = KerasRegressor(build_fn=create_basic_ANN_model, verbose=1)
    # define the grid search parameters
    batch_size = [10, 100]
    epochs = [5, 10]

    param_grid = dict(batch_size=batch_size, nb_epoch=epochs)
    grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=1, cv=7)

    grid_results =, trY)

    this works but only as a for loop over the different fp_subset, which I must define manually.
    I could easily pick the best out of every run but it wuld be great if I could fold them all inside a big grid definition and fit, so as to automatically pick the largest.

    However, until now haven’t been able to figure out a way to get that in my head.
    If the wrapper function is useful to anyone, I can post a generalised version here.

    • Jason Brownlee November 10, 2016 at 7:42 am #

      Good question.

      You might just need to us a loop around the whole lot for different projections/views of your training data.

      • L Fenu November 11, 2016 at 1:05 am #

        Thanks. I ended up coding my own for loop, saving the results of each grid in a dict, sorting the hash by the perofrmance metrics, and picking the best model.

        Now, the next question is: How do I save the model’s architecture and weights to a .json .hdf5 file? I know how to do that for a simple model. But how do I extract the best model out of the gridsearch results?

        • Jason Brownlee November 11, 2016 at 10:04 am #

          Well done.

          No need. Once you know the parameters, you can use them to train a new standalone model on all of your training data and start making predictions.

          • Fenu Luca November 15, 2016 at 3:23 am #

            I may have found a way. How about this?

            best_model = grid_result.best_estimator_.model
            best_model_file_path = ‘your_pick_here’
            model2json = best_model.to_json()
            with open( best_model_file_path+’.json’, ‘w’) as json_file:

  9. volador November 14, 2016 at 6:21 pm #

    Hi Jason, I think this is very best deep learning tutorial on the web. Thanks for your work. I have a question is :how to use the heuristic algorithm to optimize Hyperparameters for Deep Learning Models in Python With Keras, these algorithms like: Genetic algorithm, Particle swarm optimization, and Cuckoo algorithm etc. If the idea could be experimented, could you give an example

    • Jason Brownlee November 15, 2016 at 7:50 am #

      Thanks for your support volador.

      You could search the hyperparameter space using a stochastic optimization algorithm like a genetic algorithm and use the mean performance as the cost function orf fitness function. I don’t have a worked example, but it would be relatively easy to setup.

  10. Jan de Lange November 15, 2016 at 6:50 am #

    Hi Jason, very helpful intro into gridsearch for Keras. I have used your guidance in my code, but rather than using the default ‘accuracy’ to be optimized, my model requires a specific evaluation function to be optimized. You hint at this possibility in the introduction, but there is no example of it. I have followed the SciKit-learn documentation, but I fail to come up with the correct syntax.

    I have posted my question at StackOverflow, but since it is quite specific, it requires understanding of SciKit-learn in combination with Keras.

    Perhaps you can have a look? I think it would nicely extend your tutorial.

    Thanks, Jan

  11. Jan de Lange November 16, 2016 at 7:31 am #

    Yup, same sources as I referenced in my post at Stackoverflow.

  12. Anthony Ohazulike December 6, 2016 at 12:46 am #

    Good tutorial again Jason…keep on the good job!

  13. nrcjea001 December 13, 2016 at 10:48 pm #

    Hi Jason

    First off, thank you for the tutorial. It’s very helpful.

    I was also hoping you would assist on how to adapt the keras grid search to stateful lstms as discussed in

    I’ve coded the following:

    # create model
    model = KerasRegressor(build_fn=create_model, nb_epoch=1, batch_size=bats,
    verbose=2, shuffle=False)

    # define the grid search parameters
    h1n = [5, 10] # number of hidden neurons
    param_grid = dict(h1n=h1n)
    grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=5)

    for i in range(100):, trainY)

    Is grid.reset_states() corrrect? or would you suggest creating function callback for reset states.


    • Jason Brownlee December 14, 2016 at 8:27 am #

      Great question.

      With stateful LSTMs we must control the resetting of states after each epoch. The sklearn framework does not open this capacity to us – at least it looks that way to me off the cuff.

      I think you may have to grid search stateful LSTM params manually with a ton of for loops. Sorry.

      If you discover something different, let me know. i.e. there may be a way in the back door to the sklearn grid search functionality that we can inject our own custom epoch handing.

  14. Thomas Maier December 21, 2016 at 2:53 am #

    Hi Jason

    Thanks a lot for this and all the other great tutorials!

    I tried to combine this gridsearch/keras approach with a pipeline. It works if I tune nb_epoch or batch_size, but I get an error if I try to tune the optimizer or something else in the keras building function (I did not forget to include the variable as an argument):

    def keras_model(optimizer = ‘adam’):
    model = Sequential()
    model.add(Dense(80, input_dim=79, init= ‘normal’))
    model.add(Dense(1, init=’normal’))
    model.compile(optimizer=optimizer, loss=’mse’)
    return model

    kRegressor = KerasRegressor(build_fn=keras_model, nb_epoch=500, batch_size=10, verbose=0)

    estimators = []
    estimators.append((‘imputer’, preprocessing.Imputer(strategy=’mean’)))
    estimators.append((‘scaler’, preprocessing.StandardScaler()))
    estimators.append((‘kerasR’, kRegressor))
    pipeline = Pipeline(estimators)

    param_grid = dict(kerasR__optimizer = [‘adam’,’rmsprop’])

    grid = GridSearchCV(pipeline, param_grid, cv=5, scoring=’neg_mean_squared_error’)

    Do you know this problem?

    Thanks, Thomas

    • Jason Brownlee December 21, 2016 at 8:44 am #

      Thanks Thomas. I’ve not seen this issue.

      I think we’re starting to push the poor Keras sklearn wrapper to the limit.

      Maybe the next step is to build out a few functions to do manual grid searching across network configs.

  15. Jimi December 21, 2016 at 3:26 pm #

    Great resource!

    Any thoughts on how to get the “history” objects out of grid search? It could be beneficial to plot the loss and accuracy to see when a model starts to flatten out.

    • Jason Brownlee December 22, 2016 at 6:30 am #

      Not sure off the cuff Jimi, perhaps repeat the run standalone for the top performing configuration.

  16. DeepLearning January 4, 2017 at 6:08 am #

    Thanks for the post. Can we optimize the number of hidden layers as well on top of number of neurons in each layers?

    • Jason Brownlee January 4, 2017 at 9:00 am #

      Yes, it just may be very time consuming depending on the size of the dataset and the number of layers/nodes involved.

      Try it on some small datasets from the UCI ML Repo.

      • DeepLearning January 4, 2017 at 12:02 pm #

        Thanks. Would you mind looking at below code?

        def create_model(neurons=1, neurons2=1):
        # create model
        model = Sequential()
        model.add(Dense(neurons1, input_dim=8))
        model.add(Dense(1, init=’uniform’, activation=’sigmoid’))
        # Compile model
        model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])
        return model
        # define the grid search parameters
        neurons1 = [1, 3, 5, 7]
        param_grid = dict(neurons1=neurons1, neurons2=neurons2)
        grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1)
        grid_result =, Y)

        This code runs without error (I excluded certain X, y parts for brewity) but when I run “, Y), it gives AssertionError.

        I’d appreciate if you can show me where I am wrong.

        • DeepLearning January 4, 2017 at 12:26 pm #

          Update” It worked when I deleted 0 from neurons2. Thanks

        • Jason Brownlee January 5, 2017 at 9:16 am #

          A Dense() with a value of 0 neurons might blow up. Try removing the 0 from your neurons2 array.

          A good debug strategy is to cut code back to the minimum, make it work, then and add complexity. Here. Try searching a grid of 1 and 1 neurons, make it all work, then expand the grid you search.

          Let me know how you go.

  17. DeepLearning January 9, 2017 at 11:04 am #

    I keep getting error messages and I tried a big for loops that scan for all possible combinations of layer numbers, neuron numbers, other optimization stuff within defined limits. It is very time consuming code, but I could not figure it out how to adjust layer structure and other optimization parameters in the same code using GridSearch. If you would provide a code for that in your blog one day, that would be much appreciated. Thanks.

  18. Rajneesh January 11, 2017 at 10:48 am #

    Hi Jason,
    Many thanks for this awesome tutorial !

  19. Andy January 22, 2017 at 1:02 pm #

    Hi Jason,

    Great tutorial! I’m running into a slight issue. I tried running this on my own variation of the code and got the following error:

    TypeError: get_params() got an unexpected keyword argument ‘deep’

    I copied and pasted your code using the given data set and got the same error. The code is showing an error on the grid_result =, Y) line. I looked through the other comments and didn’t see anyone with the same issue. Do you know where this could be coming from?

    Thanks for your help!

    • YechiBechi January 23, 2017 at 2:18 am #

      same issue here,

      great tutorial, life saver.

    • Jason Brownlee January 23, 2017 at 8:35 am #

      Hi Andy, sorry to hear that.

      Is this happening with a specific example or with all of them?

      Are you able to check your version of Python/sklearn/keras/tf/theano?


      I can confirm the first example still works fine with Python 2.7, sklearn 0.18.1, Keras 1.2.0 and TensorFlow 0.12.1.

      • Andy January 25, 2017 at 7:12 am #

        The only differences are I am running Python 3.5 and Keras 1.2.1. The example I ran previously was the grid search for the number of neurons in a layer. But I just ran the first example and got the same error.

        Do you think the issue is due to the next version of Python? If so, what should my next steps be?

        Thanks for your help and quick response!

  20. kono February 8, 2017 at 3:14 am #


    Can you use early_stopping to decide n_epoch?

    • Jason Brownlee February 8, 2017 at 9:36 am #

      Yes, that is a good method to find a generalized model.

  21. Jayant February 23, 2017 at 4:33 am #

    Hi Jason,

    Really great article. I am a big fan of your blog and your books. Can you please explain your following statement?

    “A default cross-validation of 3 was used, but perhaps k=5 or k=10 would be more stable. Carefully choose your cross validation configuration to ensure your results are stable.”

    I didn’t see anywhere cross-validation being used.

    • Jason Brownlee February 23, 2017 at 8:56 am #

      Hi Jayant,

      Grid search uses k-fold cross-validation to evaluate the performance of each combination of parameters on unseen data.

  22. Jing February 28, 2017 at 2:09 am #

    Hi Jason,
    thanks for this awesome tutorial !
    I have two questions: 1. In “model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])”, accuracy is used for evaluate results. But GridSearchCV also has scoring parameter, if I set “scoring=’f1’”,which one is used for evaluate the results of grid search? 2.How to set two evaluate parameters ,e.g. ‘accuracy’and ’f1’ evaluating the results of grid search?

    • Jason Brownlee February 28, 2017 at 8:13 am #

      Hi Jing,

      You can set the “scoring” argument for GridSearchCV with a string of the performance measure to use, or the name of your own scoring function. You can learn about this argument here:

      You can see a full list of supported scoring measures here:

      As far as I know you can only grid search using a single measure.

      • Jing February 28, 2017 at 12:50 pm #

        Thank you so much for your help!

      • Jing February 28, 2017 at 1:54 pm #

        I find no matter what evaluate parameters used in GridSearchCV “scoring”,”metrics” in “model.compile” must be [‘accuracy’],otherwise the program gives “ValueError: The model is not configured to compute accuracy.You should pass ‘metrics=[“accuracy”]’ to the ‘model.compile()’method. So, if I set:
        model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])
        grid = GridSearchCV(estimator=model, param_grid=param_grid, scoring=’recall’)
        the grid_result.best_score_ =0.72.My question is: 0.72 is accuracy or recall ? Thank you!

        • Jason Brownlee March 1, 2017 at 8:31 am #

          Hi Jing,

          When using GridSearchCV with Keras, I would suggest not specifying any metrics when compiling your Keras model.

          I would suggest only setting the “scoring” argument on the GridSearchCV. I would expect the metric reported by GridSearchCV to be the one that you specified.

          I hope that helps.

  23. Dan March 8, 2017 at 4:13 am #

    Great Blogpost. Love it. You are awesome Jason. I got one question to GridsearchCV. As far as i understand the crossvalidation already takes place in there. That’s why we do not need any kfold anymore.
    But with this technique we would have no validation set correct? e.g. with a default value of 3 we would have 2 training sets and one test set.

    That means in kfold as well as in GridsearchCV there is no requirement for creating a validation set anymore?


    • Jason Brownlee March 8, 2017 at 9:44 am #

      Hi Dan,

      Yes, GridSearchCV performs cross validation and you must specify the number of folds. You can hold back a validation set to double check the parameters found by the search if you like. This is optional.

      • Dan March 9, 2017 at 3:25 am #

        Thank you for the quick response Jason. Especially considering the huge amount of questions you get.

  24. Johan Steunenberg March 22, 2017 at 8:25 pm #

    What I’m missing in the tutorial is the info, how to get the best params in the model with KERAS. Do I pickup the best parameters and call ‘create_model’ again with those parameters or can I call the GridSearchCV’s ‘predict’ function? (I will try out for myself but for completeness it would be good to have it in the tutorial as well.)

    • Jason Brownlee March 23, 2017 at 8:49 am #

      I see, but we don’t know the best parameters, we must search for them.

  25. Maycown Miranda April 5, 2017 at 2:09 am #

    Hi, Jason. I am getting
    /usr/local/lib/python2.7/dist-packages/keras/wrappers/ in check_params(self=, params={‘batch_size’: 10, ‘epochs’: 10})
    80 legal_params += inspect.getargspec(fn)[0]
    81 legal_params = set(legal_params)
    83 for params_name in params:
    84 if params_name not in legal_params:
    —> 85 raise ValueError(‘{} is not a legal parameter’.format(params_name))
    params_name = ‘epochs’
    87 def get_params(self, _):
    88 “””Gets parameters for this estimator.

    ValueError: epochs is not a legal parameter

    • Jason Brownlee April 9, 2017 at 2:32 pm #

      It sounds like you need to upgrade to Keras v2.0 or higher.

  26. Usman May 3, 2017 at 7:56 am #

    Nice tutorial. I would like to optimize the number of hidden layers in the model. Can you please guide in this regard, thanks

    • Jason Brownlee May 4, 2017 at 7:59 am #

      Thanks Usman.

      Consider exploring specific patterns, e.g. small-big-small, etc.

  27. Carl May 5, 2017 at 12:58 pm #

    Do you know any way this could be possible using a network with multiple inputs?

  28. DanielP May 9, 2017 at 4:26 pm #

    Hi Jason, great to see posts like this – amazing job!

    Just noticed, when you tune the optimisation algorithm SGD performs at 34% accuracy. As no parameters are being passed to the SGD function, I’d assume it takes the default configuration, lr=0.01, momentum=0.0.

    Later on, as you look for better configurations for SGD, best result (68%) is found when {‘learn_rate’: 0.01, ‘momentum’: 0.0}.

    It seems to me that these two experiments use exactly the same network configuration (including the same SGD parameters), yet their resulting accuracies differ significantly. Do you have any intuition as to why this may be happening?

  29. Pradanuari May 14, 2017 at 3:13 am #

    Hi Jason!
    absolutely love your tutorial! But would you mind to give tutorial for how to tune the number of hidden layer?


  30. Pradanuari May 14, 2017 at 11:32 pm #

    Thank you so much Jason!

  31. Ibrahim El-Fayoumi May 17, 2017 at 12:53 pm #

    Hello Jason
    I tried to use your idea in a similar problem but I am getting error : AttributeError: ‘NoneType’ object has no attribute ‘loss’
    it looks like the model does not define loss function?

    This is the error I get:
    b\site-packages\keras-2.0.4-py3.5.egg\keras\wrappers\ in fit(self=, x=memmap([[[ 0., 0., 0., …, 0., 0., 0.],
    …, 0., 0., …, 0., 0., 0.]]], dtype=float32), y=array([[ 0., 0., 0., …, 0., 0., 0.],
    [ 0., 0., 0., …, 0., 1., 0.]]), **kwargs={})
    135 self.model = self.build_fn(
    136 **self.filter_sk_params(self.build_fn.__call__))
    137 else:
    138 self.model = self.build_fn(**self.filter_sk_params(self.build_fn))
    –> 140 loss_name = self.model.loss
    loss_name = undefined
    self.model.loss = undefined
    141 if hasattr(loss_name, ‘__name__’):
    142 loss_name = loss_name.__name__
    143 if loss_name == ‘categorical_crossentropy’ and len(y.shape) != 2:
    144 y = to_categorical(y)

    AttributeError: ‘NoneType’ object has no attribute ‘loss’

    Process finished with exit code 1


    • Jason Brownlee May 18, 2017 at 8:26 am #

      Does the example in the blog post work on your system?

      • Ibrahim El-Fayoumi May 18, 2017 at 12:18 pm #

        Ok, I think your code needs to be placed after
        if __name__ == ‘__main__’:

        to work with multiprocess…

        But thanks for the post is great…

        • Jason Brownlee May 19, 2017 at 8:12 am #

          Not on Linux and OS X when I tested it, but thanks for the tip.

        • Gautam August 25, 2017 at 11:33 pm #

          n_jobs=-1 doesnt work on Windows.

          @Ibrahim: Can you please explain, what part of the code needs to be behind
          if __name__ == ‘__main__’: )

  32. Edward May 21, 2017 at 3:17 am #

    Hello Jason!
    I do the first step – try to tune Batch Size and Number of Epochs and get
    print(“Best: %f using %s” % (grid_result.best_score_, grid_result.best_params_))
    Best: 0.707031 using {‘epochs’: 100, ‘batch_size’: 40}
    After that I do the same and get
    print(“Best: %f using %s” % (grid_result.best_score_, grid_result.best_params_))
    Best: 0.688802 using {‘epochs’: 100, ‘batch_size’: 20}
    And so on
    The problem is in the grid_result.best_score_

    I expect that in the second step (for ample tuning optimizer) I will get grid_result.best_score_ better than in the first step (in the second step i use grid_result.best_params_ from the first step). But it is not true
    Tune all Hyperparameters is a very long time

    How to fix it?

    • Jason Brownlee May 21, 2017 at 6:01 am #

      Consider tuning different parameters, like network structure or number of input features.

      • Edward May 21, 2017 at 7:18 pm #

        Thanks a lot Jason!

  33. pattijane May 21, 2017 at 7:44 am #


    I’d like to have your opinion about a problem:

    I have two loss function plots, with SGD and Adamax as optimizer with same learning rate.
    Loss function of SGD looks like the red one, whereas Adamax’s looks like blue one.

    I have better scores with Adamax on validation data. I’m confused about how to proceed, should I choose Adamax and play with learning rates a little more, or go on with SGD and somehow try to improve performance?


    • Jason Brownlee May 22, 2017 at 7:49 am #

      Explore both, but focus on the validation score of interest (e.g. accuracy, RMSE, etc.) over loss.

      For example, you can get very low loss and get worse accuracy.

      • pattijane May 22, 2017 at 6:35 pm #

        Thanks for your response! I experimented with different learning rates and found out a reasonable one, (good for both Adamax and SGD) and now I try to fix learning rate and optimizer and focus on other hyperparameters such as batch-size and number of neurons. Or would be better if I set those first?

        • Jason Brownlee May 23, 2017 at 7:49 am #

          Number of neurons will have a big effect along with learning rate.

          Batch size will have a smaller effect and could be optimized last.

  34. Lotem May 23, 2017 at 1:47 am #

    Thanks for this post!

    One question – why not use grid search on all the parameters together, rather than preforming several grid searches and finding each parameter separately? surly the results are not the same…

    • Jason Brownlee May 23, 2017 at 7:54 am #

      Great question,

      In practice, the datasets are large and it can take a long time and require a lot of RAM.

  35. StatsSorceress May 25, 2017 at 6:52 am #

    Hi Jason,

    Excellent post!

    It seems to me that if you use the entire training set during your cross-validation, then your cross-validation error is going to give you an optimistically biased estimate of your validation error. I think this is because when you train the final model on the entire dataset, the validation set you create to estimate test performance comes out of the training set.

    My question is: assuming we have a lot of data, should we use perhaps only 50% of the training data for cross-validation for the hyperparameters, and then use the remaining 50% for fitting the final model (and a portion of that remaining 50% would be used for the validation set)? That way we wouldn’t be using the same data twice. I am assuming in this case that we would also have a separate test set.

    • Jason Brownlee June 2, 2017 at 11:38 am #

      Yes, it is a good idea to hold back a test set when tuning.

  36. Yang May 27, 2017 at 5:35 am #

    Thanks for your valuable post. I learned a lot from it.
    When I wrote my code for grid search, I encountered a question:

    I use fit_generator instead of fit in keras.
    Is it possible to use grid search with fit_generator ?

    I have some Merge layers in my deep learning model.
    Hence, the input of the neural network is not a single matrix.
    For example:
    Suppose we have 1,000 samples
    Input = [Input1,Input2]
    Input1 is a 1,000 *3 matrix
    Input2 is a 1,000*3*50*50 matrix (image)

    When I use the fit in your post, there is a bug….because the input1 and input2 don’t have the same dimension. So I wonder whether the fit_generator can work with grid search ?

    Thanks in advance!

  37. Yang May 27, 2017 at 6:46 am #

    Please ignore my previous reply.
    I find an answer here:
    Right now, the GridsearchCV using the scikit wrapper for network with multiple inputs is not available.

  38. Kate liu May 28, 2017 at 4:31 pm #

    Hi Jason, thank you for your good tutorial of the grid research with Keras. I followed your example with my own dataset. It could be run. But when I using the autoencoder structure, instead of the sequential structure, to gird the parameters with my own data. It could not be run. I don’t know the reason. Could you help me? Are there any differences between the gird of sequential structure and the grid of model structure?

    The follows are my codes:

    from keras.models import Sequential
    from keras.layers import Dense, Input
    from keras.wrappers.scikit_learn import KerasClassifier
    from sklearn.model_selection import StratifiedKFold
    from sklearn.model_selection import cross_val_score
    from sklearn.model_selection import GridSearchCV
    import numpy as np
    from keras.optimizers import SGD, Adam, RMSprop, Adagrad
    from keras.regularizers import l1,l2
    from keras.models import Model
    import pandas as pd
    from keras.models import load_model


    def create_model(optimizer=’rmsprop’):

    # encoder layers
    encoding_dim =140
    input_img = Input(shape=(6,))
    encoded = Dense(300, activation=’relu’,W_regularizer=l1(0.01))(input_img)
    encoded = Dense(300, activation=’relu’,W_regularizer=l1(0.01))(encoded)
    encoded = Dense(300, activation=’relu’,W_regularizer=l1(0.01))(encoded)
    encoder_output = Dense(encoding_dim, activation=’relu’,W_regularizer=l1(0.01))(encoded)

    # decoder layers
    decoded = Dense(300, activation=’relu’,W_regularizer=l1(0.01))(encoder_output)
    decoded = Dense(300, activation=’relu’,W_regularizer=l1(0.01))(decoded)
    decoded = Dense(300, activation=’relu’,W_regularizer=l1(0.01))(decoded)
    decoded = Dense(6, activation=’relu’,W_regularizer=l1(0.01))(decoded)

    # construct the autoencoder model
    autoencoder = Model(input_img, decoded)

    # construct the encoder model for plotting
    encoder = Model(input_img, encoder_output)

    # Compile model
    autoencoder.compile(optimizer=’RMSprop’, loss=’mean_squared_error’,metrics=[‘accuracy’])

    return autoencoder

    • Jason Brownlee June 2, 2017 at 12:09 pm #

      I’m surprised, I would not think the network architecture would make a difference.

      Sorry, I have no good suggestions other than try to debug the cause of the fault.

  39. Kate liu May 28, 2017 at 4:36 pm #

    the command of autoencoder.compile is modified as the follows:
    # Compile model
    autoencoder.compile(optimizer=optimizer, loss=’mean_squared_error’,metrics=[‘accuracy’])

  40. Rahul May 30, 2017 at 12:07 am #

    Can we do this for functional API as well ?

  41. Ian Worthington May 30, 2017 at 10:36 pm #

    Thanks for a great tutorial Jason, appreciated.

    njobs=-1 didn’t work very well on my Windows 10 machine: took a very long time and never finished. seems to suggest this is (or at least was in 2015) a known problem under Windows so I changed to n_jobs=1, which also allowed me to see throughput using verbose=10.

  42. Ian Worthington May 31, 2017 at 1:56 am #

    Jason —

    Given all the parameters it is possible to adjust, is there any recommendation for which should be fixed first before exploring others, or can ALL results for one change when others are changed?

  43. Mario June 9, 2017 at 12:10 am #

    Hi and thank you for the resource.

    Am I right in my understanding that this only works on one machine?

    Any hints / pointers on how to run this on a cluster? I have found as a potential avenue using Spark (no Keras though).

    Any comment at all? Information on the subject is scarce.

    • Jason Brownlee June 9, 2017 at 6:26 am #

      Yes, this example is for a single machine. Sorry, I do not have examples for running on a cluster.

  44. Shaun June 16, 2017 at 11:54 pm #

    Hi Jason,

    I’m a little bit confused about the definition of the “score” or “accuracy”. How are they made? I believe that they are not simply comparing the results with target, otherwise it will be the overfitting model being the best (like the more neurons the better).

    But on the other hand, they are just using those combinations of parameters to train the model, so what is the difference between I manually set the parameters and see my result good or not, with risk of overfitting and the grid search that creates an accuracy score to determine which one is the best?

    Best regards,

    • Jason Brownlee June 17, 2017 at 7:30 am #

      The grid search will provide an estimate of the skill of the model with a set of parameters.

      Any one configuration in the grid search can be set and evaluated manually.

      Neural networks are stochastic and will give different predictions/skill when trained on the same data.

      Ideally, if you have the time/compute the grid search should use repeated k-fold cross validation to provide robust estimates of model skill. More here:

      Does that help?

      • Shaun June 20, 2017 at 2:30 am #

        I’m new to the NN, a little bit puzzled. So say, if I have to many neurons that leads to overfitting (good on the train set, bad on the validation or test set), can grid search detect it by the score?

        My guess is yes, because there is a validation set in the GridsearchCV. Is that correct?

        • Jason Brownlee June 20, 2017 at 6:39 am #

          A larget network can overfit.

          The idea is to find a config that does well on the train and validation sets. We require a robust test harness. With enough resources, I’d recommend repeated k-fold cross validation within the grid search.

  45. Huyen June 19, 2017 at 4:21 pm #

    One more very useful tutorial, thank Jason.

    One question about GridSearch in my case. I have tried to tune parameters of my neural network for regression with 18 inputs size 800 but the time to use GridSearch totally long, like forever even though I have limited to the number. I saw in your code:

    grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1)

    Normally, n_jobs=1, can I increase that number to improve the performances?

    • Jason Brownlee June 20, 2017 at 6:36 am #

      We often cannot grid search with neural nets because it takes so long!

      Consider running on a large computer in the cloud over the weekend.

  46. Bobo June 21, 2017 at 4:57 am #

    Hi Jason,

    Any idea how to use GridSearchCV if you don’t want cross validation?

  47. makis June 28, 2017 at 11:54 pm #

    Hello. Thank you for the nice tutorial.

    I am trying to combine pipeline and gridsearch.

    Inside my keras model i use kernel_initializer=init_mode.
    Then I am trying to assign values to the init_mode dictionary in order to perform the gridsearch.

    I get the following error: ValueError: init_mode is not a legal parameter

    My code is here:

    Any tip? Thank you

  48. Abhijith Darshan Ravindra July 11, 2017 at 6:31 am #

    Hi Dr. Brownlee,

    When I run this in Spyder IDE nothing happens after

    It just appears to do nothing.

    Any suggestions as to why?

    • Jason Brownlee July 11, 2017 at 10:34 am #

      Consider running from the command line.

      The grid search may take a long time.

      • DY July 14, 2017 at 6:11 am #

        Hello Dr Brownlee,

        I saved your example codes into .py file and run it. Nothing happens after However, if I run line by line from your example codes it works. Do you know why?

        • Jason Brownlee July 14, 2017 at 8:36 am #

          It may take a long time. Consider reducing the scope of the search to see if you can get results sooner.

    • Tryfon September 18, 2017 at 11:46 pm #

      I had the same issue with you (using spyder and python 3.6) but after changing the parameter n_jobs = 1 it worked fine. Also n_jobs = 2 was stuck although spyder showed it was running in the backgound (I checked the CPU usage and was down to 1% vs the 55-80% when it is actually running).

      Don’t ask the reason why is that. My guess would be that it has to do with your system and the fact that it might not support parallelization (no CUDA GPU).

      • Jason Brownlee September 19, 2017 at 7:47 am #

        Consider running the example from the command line instead.

  49. Kamal Thapa July 27, 2017 at 3:46 pm #

    How can I do Hyper-parameter optimization for MLPRegressor in scikit learn?

  50. Josep August 3, 2017 at 2:31 am #

    Hi Jason,
    I’m unable to apply the grid search to a seq to seq LSTM network (Keras Regressor model in the scikit API). When I set the GridSearchCV scoring algorithm to r^2 (or any scoring function for regression problems) the expect a 2 dim input vector, not the 3 dim used in Keras.
    Otherwise, if I left the default scoring algorithm named “_passthrough_scorer”( I don’t know what it does, I don’t even know what it is) it works but the best_score doesn’t match with the real best parametrization. I’m really confused…I’ll had to write the grid search manually…

    • Josep August 3, 2017 at 2:42 am #

      I’ve solved it, I share it if someone have the same issue…,If you set the gridsearch scoring function to “None” it uses the scoring metrics of the Keras model.

      • Josep August 3, 2017 at 2:49 am #

        Sorry for bothering, but the results of the approach I’ve said are incorrect. I don’t know what to do.

    • Jason Brownlee August 3, 2017 at 6:54 am #

      Hi Josep,

      Consider writing your own for loop to iterate over params and run a Cross Validation for the params within the loop.

      This is how I do it now for large/complex models.

  51. kotb August 8, 2017 at 7:10 pm #

    Can i use this grid search without using keras model

  52. Aman Garg August 19, 2017 at 3:35 am #

    Hello Jason,

    Thanks for such a nice tutorial.

    Instead of getting a output as ‘Best: 0.720052 using {‘init_mode’: ‘uniform’}’ , it would be really nice if you could show us how to visualize this result with matplotlib so that it gets more easier.

  53. Michael August 20, 2017 at 4:42 am #

    Hi, Jason. Thanks, again, for all of the blog posts and example code. I’m trying to tune my binary classification Keras neural network. My dataset includes about 50,000 entries with 52 (numeric) variables. Using Grid Search, I’ve tested all sorts of combinations of layer size, number of epochs, batch size, optimizers, activations, learning rates, dropout rates, and L2 regularization parameters. My grid search shows every combination performs the same. For example, here is a snippet from my latest results:

    Best: 0.876381 using {‘act’: ‘relu’, ‘opt’: ‘Adam’}
    0.876381 (0.003878) with: {‘act’: ‘relu’, ‘opt’: ‘Adam’}
    0.876381 (0.003878) with: {‘act’: ‘relu’, ‘opt’: ‘SGD’}
    0.876381 (0.003878) with: {‘act’: ‘relu’, ‘opt’: ‘Adagrad’}
    0.876381 (0.003878) with: {‘act’: ‘relu’, ‘opt’: ‘Adadelta’}
    0.876361 (0.003880) with: {‘act’: ‘tanh’, ‘opt’: ‘Adam’}
    0.876381 (0.003878) with: {‘act’: ‘tanh’, ‘opt’: ‘SGD’}

    But I also get 0.876381 whether I have 1000 nodes or 1 node, and for every other combo I’ve tested. I’ve also tried different ways of scaling or transforming my input data with no impact.

    Do you have any thoughts on why I’m having trouble finding different combinations of parameters that actually have a difference in performance?

    Thank you for your help! You rock!

  54. Shubham Kumar September 3, 2017 at 11:54 am #

    Hey Jason.
    I was using grid search to tune hyperparameters for a CNN-LSTM classification problem.
    I used the code template on your blog about sequence classification.
    MY original data has 38932 instances, but for tuning I am using only 1000 to save time.
    But even then, I am not sure how to best search for those parameters and save time.

    Is it a bad idea to search for hyper parameters in a small subset (almost 1/40th of training in my case).
    Will the result vary largely when I use actual data size?
    Also, I passed in several parameters for the grid search. Left it overnight and it still hadn’t made enough progress, so I stopped the execution.
    How can I speed up this process?

    • Jason Brownlee September 3, 2017 at 3:44 pm #

      The result will be biased, but perhaps might give you an idea of the direction in which to proceed – this could be enough for you.

      I often run a lot of sanity check grid searches on small samples to get ideas on which direction to push.

      More data will result in less biased estimates of model skill, often proportionately to some point of diminishing returns.

      • Shubham Kumar September 4, 2017 at 3:10 am #

        Great !
        I did read that one of the sanity checks is to check whether the model overfits on a small sample! If yes, then we are good to go…
        I am slightly new to building proper models and find this part exciting but a little intimidating at the same time !
        I am going to use only a few hyper parameters at a time, and keep the rest constant and check what happens !

        Love your posts ! They are amazingly helpful .
        Does the Python LSTM book have code snippets in Python 3 as well?
        Coz it becomes a little difficult to search for the right modules and attributes otherwise :/

        • Jason Brownlee September 4, 2017 at 4:39 am #


          Yes, the code in my LSTM book was tested with Python 2.7 and Python 3.5.

  55. Kaushal Shetty September 8, 2017 at 12:24 am #

    Hi Jason, Is this a valid approach to decide the number of layers?
    def neural_train(layer1 = 1,layer2 = 1,layer3 = 1,layers = 1):

    input_tensor = Input(shape=(2001,))
    x = Dense(units = layer1,activation=’relu’)(input_tensor)
    if layers == 2:
    x = Dense(layer2,activation = ‘relu’)(x)
    if layers ==3 :
    x = Dense(layer2,activation = ‘relu’)(x)
    x = Dense(layer3,activation = ‘relu’)(x)

    output_tensor = Dense(10,activation=’softmax’)(x)
    model = Model(input_tensor,output_tensor)
    model.compile(optimizer = ‘rmsprop’,loss=’categorical_crossentropy’,metrics = [‘accuracy’])
    return model

    layer1 = [1024,512]
    layer2 = [256,100]
    layer3 = [60,40]
    epochs = [10,11]
    layers = [2,3]
    param_grid = dict(epochs = epochs,layer1 = layer1,layer2 = layer2,layer3 = layer3,layers=layers)
    model = KerasClassifier(build_fn = neural_train)
    gsv_model = GridSearchCV(model,param_grid=param_grid),y_train)

  56. ari September 9, 2017 at 1:29 am #

    Very helpful post Jason. Thanks for this. Are there any advantages for using gridsearch over something like hyperas/hyperopt ? To your best knowledge is one faster than the other?

    • Jason Brownlee September 9, 2017 at 11:58 am #

      Depends on your data and model. Use the took that you prefer.

  57. Shubham Kumar September 10, 2017 at 4:38 am #

    {‘split0_test_score’: array([ 0.6641791, 0.6641791, 0.6641791, 0.6641791]), ‘split1_test_score’: array([ 0.65413534, 0.65413534, 0.65413534, 0.65413534]), ‘split2_test_score’: array([ 0.69924811, 0.69924811, 0.69924811, 0.69924811]), ‘mean_test_score’: array([ 0.6725, 0.6725, 0.6725, 0.6725]), ‘std_test_score’: array([ 0.01931902, 0.01931902, 0.01931902, 0.01931902]), ‘rank_test_score’: array([1, 1, 1, 1]), ‘split0_train_score’: array([ 0.67669174, 0.67669174, 0.67669174, 0.67669174]), ‘split1_train_score’: array([ 0.68164794, 0.68164794, 0.68164794, 0.68164794]), ‘split2_train_score’: array([ 0.65917602, 0.65917602, 0.65917602, 0.65917602]), ‘mean_train_score’: array([ 0.67250523, 0.67250523, 0.67250523, 0.67250523]), ‘std_train_score’: array([ 0.00963991, 0.00963991, 0.00963991, 0.00963991]), ‘mean_fit_time’: array([ 36.72573058, 37.0244147 , 38.12670692, 40.71116368]), ‘std_fit_time’: array([ 0.4829061 , 0.35207924, 0.13746276, 2.71443639]), ‘mean_score_time’: array([ 1.49508754, 1.76741695, 2.14029002, 2.67426189]), ‘std_score_time’: array([ 0.04907801, 0.11919153, 0.07953362, 0.13931651]), ‘param_dropout’: masked_array(data = [0.2 0.5 0.6 0.7],
    mask = [False False False False],
    fill_value = ?)
    , ‘params’: ({‘dropout’: 0.2}, {‘dropout’: 0.5}, {‘dropout’: 0.6}, {‘dropout’: 0.7})}

    Hey. I was hypertuning a model on 4 different choices of hyper parameters. However, in the grid_results_ dictionary, the rank_test_score key has array with all same values. I find that confusing. Shouldn’t it have 4 different values in each place?
    Something like [1,3,2,4] ?
    What could be the explanation for this?

    • Shubham Kumar September 10, 2017 at 4:50 am #

      It must have something to do with all mean_test_scores being the same ,

    • Jason Brownlee September 11, 2017 at 12:03 pm #

      If you are testing 4 different values for one parameter, then you must build 4 models/complete 4 runs.

      Does that help?

      • Shubham Kumar September 13, 2017 at 5:20 am #

        I am sorry. That’s confusing. 4 models or complete 4 runs means ?

        Things are different if we are gridsearching/randomsearching just for one hyperparameter?

        Does it have something to do with the actual code used to write TensorFlow /keras ?

        • Jason Brownlee September 13, 2017 at 12:36 pm #

          If you have one parameter and you want to test 4 values, each value needs one run. Ideally, we would run many times for each parameter value and take the average skill score given the stochastic nature of ML algorithms.

          For a random search, you run for as long as you like.

          Does that help?

  58. Shubham Kumar September 13, 2017 at 11:17 pm #

    What I understand is that when we have more than 1 (say 2) hyper-parameters in a grid, then for each combination, the code will complete as many epochs as I have specified, with as many training-cross-validation sets as specified (the CV in GridSearchCV). So, going through all those epochs, for each training-cross-validation set, we get the avg accuracy over all the cross-validation sets for every combination.

    So when you say 1 run only in the case of a single hyperparameter, that means only 1 training-crossvalidation set? Because only in this case, there won’t be any averaging involved.

    Is that what I have to do? Change the training-crossValidation set to just 1?

  59. Rishi September 18, 2017 at 5:18 am #

    would you please post an example of inheriting from KerasClassifier (or KerasRegressor) to create your own class? I’m attempting to do this and it works for the most part:

    class MLP_Regressor(KerasRegressor):

    def __init__(self, **sk_params):
    super().__init__(build_fn=None, **sk_params)

    def __call__(self, optimizer=’adam’, loss=’mean_squared_error’, **kwargs):
    # more code goes here (that was previously in ‘build_fn’

    I can include this in a pipeline and it runs perfectly:
    MLP Pipeline(memory=None,
    steps=[(‘MLP’, )])

    Only thing is: The Keras documentation includes the ‘build_fn’ keyword argument:

    keras.wrappers.scikit_learn.KerasClassifier(build_fn=None, **sk_params)

    While the actual KerasClassifier class definition shows the following in its __init__ method:

    def __init__(self, model, optimizer=’adam’, loss=’categorical_crossentropy’, **kwargs):
    super(KerasClassifier, self).__init__(model, optimizer, loss, **kwargs)

    I’m not sure if my __init__ in MLP_Regressor has been setup correctly (to avoid hidden bugs in the future).

    Would greatly appreciate it! (I’ve searched, but couldn’t find a single example of KerasClassifier inheritance).

    • Jason Brownlee September 18, 2017 at 5:49 am #

      Thanks for the suggestion, I have not done this but perhaps in the future.

  60. Tmn September 20, 2017 at 2:45 am #

    Hi Jason,

    I can not thank you enough. I am sure that there are many people like me who have learnt a lost from your tutorial on both “R” and “Python”. I have been following your tutorial for more than 3 year now. Before I was using R however, recently I moved to python for Deep learning. And I find your tutorial as usual, exceptional. I think Andrew Ng and CS231n (andrej karpathy), theoretical course and your programming course on deep learning is one of the best in the world. You rock! Thanks a lot.

    I do have a question 🙂 as well.
    The grid search parameter tuning works perfectly with CPU. I agree with your suggestion not to tune everything at once. Now I moved to GPU implementation. I was able to execute the code if I chose options n_job=1. However, if I do multi-threading n_job=-1. I am getting “CUDA_ERROR_OUT_OF_MEMORY”. I have GeForce GTX 1080. Did you happen to encounter similar kind of error? I will post you the error log if needed.

    Once, again thank you.

    • Jason Brownlee September 20, 2017 at 6:00 am #

      Thanks for all of your support!

      Yes, I have the same and I would recommend using a “single thread” and let the GPU do its thing for a given single run.

      In general, I’d recommend contrasting different approaches to grid searching (cpu/gpu) and use the approach that is overall faster for your specific tests.

      • Tmn September 20, 2017 at 11:33 pm #

        Hi Jason,
        Thank you for the response. The parameter search using CPU (n_job=-1) is (2.961489-4.977758) while using GPU (n_job=1) is (140.101048-142.151023) second.

        One more thing, after grid search I have value for parameters {batch_size, activation, neurons, learn_rate..} and accuracy around 90%. However, I wonder why reusing these model parameter does not provide the same results, now accuracy is 52%. Even though I executed it many times with same parameter the accuracy remains the same (52%). I could not achieve the accuracy as shown in grid search using best model parameter. I am doing 5-fold CV I do not expect the accuracy to be the same since it is stochastic process but it should be around SD±5%. What do you think? Did you also happen to encounter the same thing ?

        Also the best parameter values changes in each executions with an accuracy SD±5%.


        Below code is something I am doing to limit GPU memory usage and run multiple grid search. However, we should know the memory usage in advance ( Let me know if it makes sense.

        Also, we can use n-job. I tried with n_job = 2 however the GPU memory is allocated based on fraction. I am searching how to allocated memory based on MB. I will do more research on this “CUDA_ERROR_OUT_OF_MEMORY” and update you.

        import tensorflow as tf
        from keras.backend.tensorflow_backend import set_session
        config = tf.ConfigProto()
        config.gpu_options.per_process_gpu_memory_fraction = 0.3


        • Jason Brownlee September 21, 2017 at 5:42 am #

          The results for the standalone model should fit into the distribution of the grid search results – if you repeated each grid search result many times, e.g. 10-30. See this post on evaluating model skill of neural networks:

          Nice, sorry, I cannot give you good advice on grid searching with the GPU, it is not something I do generally. I am more likely to run instances serially or across AWS instances.

          • TMN October 6, 2017 at 2:12 am #

            Hi Jason,

            Could you please help on how to do features normalization while doing the grid search and cross-validation. Is normalization is done automatically here, GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=15,cv=rkf)? If I normalize the features during training X = scaler.transform(X_train), this will introduce bias in cross-validation. Also, if possible, can you please provide me references on using scikit-learn wrapper with Keras for advance options, is their any limitation on wrapper ?

            Without normalization:
            Best: 0.535211 using {‘learn_rate’: 0.01, ‘dropout_rate’: 25, ‘batch_size’: 40, ‘neurons’: 200, ‘init_mode’: ‘lecun_uniform’, ‘optimizer’: ‘SGD’, ‘activation’: ‘relu’, ‘epochs’: 1000}

            With normalization:
            Best: 0.695775 using {‘optimizer’: ‘SGD’, ‘batch_size’: 132, ‘init_mode’: ‘lecun_uniform’, ‘epochs’: 1000, ‘learn_rate’: 0.01, ‘dropout_rate’: 25, ‘neurons’: 200, ‘activation’: ‘relu’}

          • Jason Brownlee October 6, 2017 at 5:37 am #

            Perhaps you can normalize your data prior to the grid search?

          • TMN October 6, 2017 at 10:59 am #

            I normalize my data prior to grid search using X = scaler.transform(X_train) but dont you think it would introduce bias in the performance. Normally, I expect to normalize train set and use that normalization factor to normalize test or validation set before prediction. May be I did not understand you properly, how do you do normalization prior to grid search?


          • Jason Brownlee October 6, 2017 at 11:07 am #

            Yes, it’s a struggle or trade-off.

            Perhaps you can see if a Pipeline will work in the grid search, it may, but I expect it will error.

            Perhaps the bias is minor and you can ignore it.

            Perhaps you can implement your own grid search loop to only use training data to calculate data scaling coefficients.

          • TMN October 6, 2017 at 6:44 pm #

            I started looking at the pipeline ( on how they have been using it for SVM, lets see. I would expect the pipeline to work for Keras as well, as this is a classical problem in machine learning. Why do you expect error here? I wanted to take the full advantage from automatic grid search. Well, the final option will be to implement my own grid search.

            The bias is really significant in 5-repeated 10-fold CV. Thanks

            Without normalization:
            Best: 0.535211 using {‘learn_rate’: 0.01, ‘dropout_rate’: 25, ‘batch_size’: 40, ‘neurons’: 200, ‘init_mode’: ‘lecun_uniform’, ‘optimizer’: ‘SGD’, ‘activation’: ‘relu’, ‘epochs’: 1000}

            With normalization:
            Best: 0.695775 using {‘optimizer’: ‘SGD’, ‘batch_size’: 132, ‘init_mode’: ‘lecun_uniform’, ‘epochs’: 1000, ‘learn_rate’: 0.01, ‘dropout_rate’: 25, ‘neurons’: 200, ‘activation’: ‘relu’}

          • Jason Brownlee October 7, 2017 at 5:51 am #

            If it works, that is great. I have seen cases where when grid search + keras gets fancy it causes errors.

            I have a tutorial on Pipeline here that might help:

  61. HWU September 22, 2017 at 6:52 am #

    This is such a great, thorough tutorial. Thanks for keeping your tutorials up to date! It’s so nice finding a resource with examples that you know will work because they’ve been tested on recent versions of required packages.

  62. Marjan September 29, 2017 at 1:08 pm #

    Thank you for your great tutorial. I tried to use it for my model with multiple inputs. but It didn`t work. I found that the scikit-learn wrapper does not work for multiple inputs. it gives me an error for[input1,input2],y)
    Do you have any suggestion to handle it?

    • Jason Brownlee September 30, 2017 at 7:34 am #

      Sorry I do not. Perhaps run the grid search manually (e.g. your own for loop)?

  63. Buz Fifer October 5, 2017 at 7:06 am #

    When I run your code to tune the dropout_rate, I get the following error:
    ValueError: dropout_rate is not a legal parameter

    In fact, I get this error for all labels except epochs and batch_size. Both of these were recognized and ran fine. I could not find a reference to valid labels anywhere, even in API docs. Any suggestions?

    • Jason Brownlee October 5, 2017 at 5:16 pm #

      What do you mean by valid labels exactly?

      • Buz Fifer October 6, 2017 at 3:02 am #

        Sorry, I should have included the code in the first place. I have added comments in the code to show exactly what I tried for each parameter.

        # ———— Define Keras Classifier Wrapper
        model1 = KerasClassifier(build_fn=kerasModel1, epochs=5, batch_size=10, verbose=0)

        # ———– define the grid search parameters
        mybatchs = [10, 20, 128]
        myepochs = [5, 10, 20, 50, 60, 80, 100]
        mylearn = [0.001, 0.002, 0.0025, 0.003]
        myopts = [‘Adam’, ‘Nadam’, ‘RMSprop’]
        myinits = [‘uniform’, ‘normal’, ‘lecun_uniform’, ‘lecun_normal’, ‘glorot_uniform’, ‘glorot_normal’]
        mydrop = [0.10, 0.20, 0.30, 0.35, 0.40, 0.50, 0.60, 0.70, 0.80]

        # ————- Not Recognized
        #param_grid = dict(optimizer=myopts)
        #param_grid = dict(learn_rate=mylearn)
        #param_grid = dict(learning_rate=mylearn)
        #param_grid = dict(init=myinits)
        #param_grid = dict(init_mode=myinits)
        #param_grid = dict(dropout_rate=mydrop)

        # ———— Recognized
        #param_grid = dict(epochs=myepochs) # —– OK
        #param_grid = dict(batch_size=mybatchs) # —– OK

        I removed comment # and ran each one separately. For example, running the first param_grid values resulted in: Error – optimizer is not a valid parameter. They all got the same rejection notice except for epochs and batch_size.
        I hope that helps.

  64. Buz Fifer October 6, 2017 at 3:09 am #

    Just to be clearer, each parameter had it’s own name in the error message as follows:

    Error – optimizer is not a valid parameter
    Error – learn_rate is not a valid parameter
    Error – learning_rate is not a valid parameter
    Error – init is not a valid parameter
    Error – init_mode is not a valid parameter
    Error – dropout_rate is not a valid parameter

    • Jason Brownlee October 6, 2017 at 5:39 am #

      That is odd, I don’t have any good ideas, other than continue to debug and try different variations to see if you can expose the cause of the issue.

      Double check all of your python libraries are up to date.

  65. ritika October 6, 2017 at 11:49 pm #

    Hi Jason, Very nice tutorial..very well explained

  66. TC October 17, 2017 at 10:27 am #

    Hi Jason thanks for the great post.

    Let’s say I’m using 5 fold CV on a relatively small dataset (not necessarily for a deep learning model). In this case, the variance of the performance metric might be quite high, and just by chance, a point on the grid that is in reality far from optimal, might be selected as the “best”.

    So are there any approaches to smooth out the response surface of the grid search, to deal with “spikes” in performance due to variance?

    • Jason Brownlee October 17, 2017 at 4:05 pm #

      Wonderful question.

      Yes, we can approach this problem by increasing the number of repeats (not folds) of each param combination.

Leave a Reply