Last Updated on August 27, 2020

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.

**Kick-start your project** with my new book Deep Learning With Python, including *step-by-step tutorials* and the *Python source code* files for all examples.

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”.**Update March/2018**: Added alternate link to download the dataset.**Update Oct/2019**: Updated for Keras 2.3.0 API.

## Overview

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:

- How to use Keras models in scikit-learn.
- How to use grid search in scikit-learn.
- How to tune batch size and training epochs.
- How to tune optimization algorithms.
- How to tune learning rate and momentum.
- How to tune network weight initialization.
- How to tune activation functions.
- How to tune dropout regularization.
- 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:

1 2 3 4 5 |
def create_model(): ... return model model = KerasClassifier(build_fn=create_model) |

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

For example:

1 2 3 4 5 |
def create_model(): ... return model model = KerasClassifier(build_fn=create_model, epochs=10) |

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:

1 2 3 4 5 |
def create_model(dropout_rate=0.0): ... return model model = KerasClassifier(build_fn=create_model, dropout_rate=0.2) |

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:

1 2 3 |
param_grid = dict(epochs=[10,20,30]) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) |

Once completed, you can access the outcome of the grid search in the result object returned from **grid.fit()**. 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.

- Download the dataset and place it in your currently working directly with the name
**pima-indians-diabetes.csv**(update: download from here).

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:

1 2 |
INFO (theano.gof.compilelock): Waiting for existing lock by process '55614' (I am process '55613') INFO (theano.gof.compilelock): To manually release the lock, delete ... |

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 code).

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

## 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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 |
# Use scikit-learn to grid search the batch size and epochs import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.wrappers.scikit_learn import KerasClassifier # Function to create model, required for KerasClassifier def create_model(): # create model model = Sequential() model.add(Dense(12, input_dim=8, activation='relu')) model.add(Dense(1, activation='sigmoid')) # Compile model model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, verbose=0) # define the grid search parameters batch_size = [10, 20, 40, 60, 80, 100] epochs = [10, 50, 100] param_grid = dict(batch_size=batch_size, epochs=epochs) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 |
Best: 0.686198 using {'epochs': 100, 'batch_size': 20} 0.348958 (0.024774) with: {'epochs': 10, 'batch_size': 10} 0.348958 (0.024774) with: {'epochs': 50, 'batch_size': 10} 0.466146 (0.149269) with: {'epochs': 100, 'batch_size': 10} 0.647135 (0.021236) with: {'epochs': 10, 'batch_size': 20} 0.660156 (0.014616) with: {'epochs': 50, 'batch_size': 20} 0.686198 (0.024774) with: {'epochs': 100, 'batch_size': 20} 0.489583 (0.075566) with: {'epochs': 10, 'batch_size': 40} 0.652344 (0.019918) with: {'epochs': 50, 'batch_size': 40} 0.654948 (0.027866) with: {'epochs': 100, 'batch_size': 40} 0.518229 (0.032264) with: {'epochs': 10, 'batch_size': 60} 0.605469 (0.052213) with: {'epochs': 50, 'batch_size': 60} 0.665365 (0.004872) with: {'epochs': 100, 'batch_size': 60} 0.537760 (0.143537) with: {'epochs': 10, 'batch_size': 80} 0.591146 (0.094954) with: {'epochs': 50, 'batch_size': 80} 0.658854 (0.054904) with: {'epochs': 100, 'batch_size': 80} 0.402344 (0.107735) with: {'epochs': 10, 'batch_size': 100} 0.652344 (0.033299) with: {'epochs': 50, 'batch_size': 100} 0.542969 (0.157934) with: {'epochs': 100, 'batch_size': 100} |

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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 |
# Use scikit-learn to grid search the batch size and epochs import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.wrappers.scikit_learn import KerasClassifier # Function to create model, required for KerasClassifier def create_model(optimizer='adam'): # create model model = Sequential() model.add(Dense(12, input_dim=8, activation='relu')) model.add(Dense(1, activation='sigmoid')) # Compile model model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, verbose=0) # define the grid search parameters optimizer = ['SGD', 'RMSprop', 'Adagrad', 'Adadelta', 'Adam', 'Adamax', 'Nadam'] param_grid = dict(optimizer=optimizer) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 |
Best: 0.704427 using {'optimizer': 'Adam'} 0.348958 (0.024774) with: {'optimizer': 'SGD'} 0.348958 (0.024774) with: {'optimizer': 'RMSprop'} 0.471354 (0.156586) with: {'optimizer': 'Adagrad'} 0.669271 (0.029635) with: {'optimizer': 'Adadelta'} 0.704427 (0.031466) with: {'optimizer': 'Adam'} 0.682292 (0.016367) with: {'optimizer': 'Adamax'} 0.703125 (0.003189) with: {'optimizer': 'Nadam'} |

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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 |
# Use scikit-learn to grid search the learning rate and momentum import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.wrappers.scikit_learn import KerasClassifier from keras.optimizers import SGD # Function to create model, required for KerasClassifier def create_model(learn_rate=0.01, momentum=0): # create model model = Sequential() model.add(Dense(12, input_dim=8, activation='relu')) model.add(Dense(1, activation='sigmoid')) # Compile model optimizer = SGD(lr=learn_rate, momentum=momentum) model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, verbose=0) # define the grid search parameters learn_rate = [0.001, 0.01, 0.1, 0.2, 0.3] momentum = [0.0, 0.2, 0.4, 0.6, 0.8, 0.9] param_grid = dict(learn_rate=learn_rate, momentum=momentum) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 |
Best: 0.680990 using {'learn_rate': 0.01, 'momentum': 0.0} 0.348958 (0.024774) with: {'learn_rate': 0.001, 'momentum': 0.0} 0.348958 (0.024774) with: {'learn_rate': 0.001, 'momentum': 0.2} 0.467448 (0.151098) with: {'learn_rate': 0.001, 'momentum': 0.4} 0.662760 (0.012075) with: {'learn_rate': 0.001, 'momentum': 0.6} 0.669271 (0.030647) with: {'learn_rate': 0.001, 'momentum': 0.8} 0.666667 (0.035564) with: {'learn_rate': 0.001, 'momentum': 0.9} 0.680990 (0.024360) with: {'learn_rate': 0.01, 'momentum': 0.0} 0.677083 (0.026557) with: {'learn_rate': 0.01, 'momentum': 0.2} 0.427083 (0.134575) with: {'learn_rate': 0.01, 'momentum': 0.4} 0.427083 (0.134575) with: {'learn_rate': 0.01, 'momentum': 0.6} 0.544271 (0.146518) with: {'learn_rate': 0.01, 'momentum': 0.8} 0.651042 (0.024774) with: {'learn_rate': 0.01, 'momentum': 0.9} 0.651042 (0.024774) with: {'learn_rate': 0.1, 'momentum': 0.0} 0.651042 (0.024774) with: {'learn_rate': 0.1, 'momentum': 0.2} 0.572917 (0.134575) with: {'learn_rate': 0.1, 'momentum': 0.4} 0.572917 (0.134575) with: {'learn_rate': 0.1, 'momentum': 0.6} 0.651042 (0.024774) with: {'learn_rate': 0.1, 'momentum': 0.8} 0.651042 (0.024774) with: {'learn_rate': 0.1, 'momentum': 0.9} 0.533854 (0.149269) with: {'learn_rate': 0.2, 'momentum': 0.0} 0.427083 (0.134575) with: {'learn_rate': 0.2, 'momentum': 0.2} 0.427083 (0.134575) with: {'learn_rate': 0.2, 'momentum': 0.4} 0.651042 (0.024774) with: {'learn_rate': 0.2, 'momentum': 0.6} 0.651042 (0.024774) with: {'learn_rate': 0.2, 'momentum': 0.8} 0.651042 (0.024774) with: {'learn_rate': 0.2, 'momentum': 0.9} 0.455729 (0.146518) with: {'learn_rate': 0.3, 'momentum': 0.0} 0.455729 (0.146518) with: {'learn_rate': 0.3, 'momentum': 0.2} 0.455729 (0.146518) with: {'learn_rate': 0.3, 'momentum': 0.4} 0.348958 (0.024774) with: {'learn_rate': 0.3, 'momentum': 0.6} 0.348958 (0.024774) with: {'learn_rate': 0.3, 'momentum': 0.8} 0.348958 (0.024774) with: {'learn_rate': 0.3, 'momentum': 0.9} |

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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 |
# Use scikit-learn to grid search the weight initialization import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.wrappers.scikit_learn import KerasClassifier # Function to create model, required for KerasClassifier def create_model(init_mode='uniform'): # create model model = Sequential() model.add(Dense(12, input_dim=8, kernel_initializer=init_mode, activation='relu')) model.add(Dense(1, kernel_initializer=init_mode, activation='sigmoid')) # Compile model model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, verbose=0) # define the grid search parameters init_mode = ['uniform', 'lecun_uniform', 'normal', 'zero', 'glorot_normal', 'glorot_uniform', 'he_normal', 'he_uniform'] param_grid = dict(init_mode=init_mode) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 9 |
Best: 0.720052 using {'init_mode': 'uniform'} 0.720052 (0.024360) with: {'init_mode': 'uniform'} 0.348958 (0.024774) with: {'init_mode': 'lecun_uniform'} 0.712240 (0.012075) with: {'init_mode': 'normal'} 0.651042 (0.024774) with: {'init_mode': 'zero'} 0.700521 (0.010253) with: {'init_mode': 'glorot_normal'} 0.674479 (0.011201) with: {'init_mode': 'glorot_uniform'} 0.661458 (0.028940) with: {'init_mode': 'he_normal'} 0.678385 (0.004872) with: {'init_mode': 'he_uniform'} |

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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 |
# Use scikit-learn to grid search the activation function import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.wrappers.scikit_learn import KerasClassifier # Function to create model, required for KerasClassifier def create_model(activation='relu'): # create model model = Sequential() model.add(Dense(12, input_dim=8, kernel_initializer='uniform', activation=activation)) model.add(Dense(1, kernel_initializer='uniform', activation='sigmoid')) # Compile model model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, verbose=0) # define the grid search parameters activation = ['softmax', 'softplus', 'softsign', 'relu', 'tanh', 'sigmoid', 'hard_sigmoid', 'linear'] param_grid = dict(activation=activation) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 9 |
Best: 0.722656 using {'activation': 'linear'} 0.649740 (0.009744) with: {'activation': 'softmax'} 0.720052 (0.032106) with: {'activation': 'softplus'} 0.688802 (0.019225) with: {'activation': 'softsign'} 0.720052 (0.018136) with: {'activation': 'relu'} 0.691406 (0.019401) with: {'activation': 'tanh'} 0.680990 (0.009207) with: {'activation': 'sigmoid'} 0.691406 (0.014616) with: {'activation': 'hard_sigmoid'} 0.722656 (0.003189) with: {'activation': 'linear'} |

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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 |
# Use scikit-learn to grid search the dropout rate import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.layers import Dropout from keras.wrappers.scikit_learn import KerasClassifier from keras.constraints import maxnorm # Function to create model, required for KerasClassifier def create_model(dropout_rate=0.0, weight_constraint=0): # create model model = Sequential() model.add(Dense(12, input_dim=8, kernel_initializer='uniform', activation='linear', kernel_constraint=maxnorm(weight_constraint))) model.add(Dropout(dropout_rate)) model.add(Dense(1, kernel_initializer='uniform', activation='sigmoid')) # Compile model model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, verbose=0) # define the grid search parameters weight_constraint = [1, 2, 3, 4, 5] dropout_rate = [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] param_grid = dict(dropout_rate=dropout_rate, weight_constraint=weight_constraint) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 |
Best: 0.723958 using {'dropout_rate': 0.2, 'weight_constraint': 4} 0.696615 (0.031948) with: {'dropout_rate': 0.0, 'weight_constraint': 1} 0.696615 (0.031948) with: {'dropout_rate': 0.0, 'weight_constraint': 2} 0.691406 (0.026107) with: {'dropout_rate': 0.0, 'weight_constraint': 3} 0.708333 (0.009744) with: {'dropout_rate': 0.0, 'weight_constraint': 4} 0.708333 (0.009744) with: {'dropout_rate': 0.0, 'weight_constraint': 5} 0.710937 (0.008438) with: {'dropout_rate': 0.1, 'weight_constraint': 1} 0.709635 (0.007366) with: {'dropout_rate': 0.1, 'weight_constraint': 2} 0.709635 (0.007366) with: {'dropout_rate': 0.1, 'weight_constraint': 3} 0.695312 (0.012758) with: {'dropout_rate': 0.1, 'weight_constraint': 4} 0.695312 (0.012758) with: {'dropout_rate': 0.1, 'weight_constraint': 5} 0.701823 (0.017566) with: {'dropout_rate': 0.2, 'weight_constraint': 1} 0.710938 (0.009568) with: {'dropout_rate': 0.2, 'weight_constraint': 2} 0.710938 (0.009568) with: {'dropout_rate': 0.2, 'weight_constraint': 3} 0.723958 (0.027126) with: {'dropout_rate': 0.2, 'weight_constraint': 4} 0.718750 (0.030425) with: {'dropout_rate': 0.2, 'weight_constraint': 5} 0.721354 (0.032734) with: {'dropout_rate': 0.3, 'weight_constraint': 1} 0.707031 (0.036782) with: {'dropout_rate': 0.3, 'weight_constraint': 2} 0.707031 (0.036782) with: {'dropout_rate': 0.3, 'weight_constraint': 3} 0.694010 (0.019225) with: {'dropout_rate': 0.3, 'weight_constraint': 4} 0.709635 (0.006639) with: {'dropout_rate': 0.3, 'weight_constraint': 5} 0.704427 (0.008027) with: {'dropout_rate': 0.4, 'weight_constraint': 1} 0.717448 (0.031304) with: {'dropout_rate': 0.4, 'weight_constraint': 2} 0.718750 (0.030425) with: {'dropout_rate': 0.4, 'weight_constraint': 3} 0.718750 (0.030425) with: {'dropout_rate': 0.4, 'weight_constraint': 4} 0.722656 (0.029232) with: {'dropout_rate': 0.4, 'weight_constraint': 5} 0.720052 (0.028940) with: {'dropout_rate': 0.5, 'weight_constraint': 1} 0.703125 (0.009568) with: {'dropout_rate': 0.5, 'weight_constraint': 2} 0.716146 (0.029635) with: {'dropout_rate': 0.5, 'weight_constraint': 3} 0.709635 (0.008027) with: {'dropout_rate': 0.5, 'weight_constraint': 4} 0.703125 (0.011500) with: {'dropout_rate': 0.5, 'weight_constraint': 5} 0.707031 (0.017758) with: {'dropout_rate': 0.6, 'weight_constraint': 1} 0.701823 (0.018688) with: {'dropout_rate': 0.6, 'weight_constraint': 2} 0.701823 (0.018688) with: {'dropout_rate': 0.6, 'weight_constraint': 3} 0.690104 (0.027498) with: {'dropout_rate': 0.6, 'weight_constraint': 4} 0.695313 (0.022326) with: {'dropout_rate': 0.6, 'weight_constraint': 5} 0.697917 (0.014382) with: {'dropout_rate': 0.7, 'weight_constraint': 1} 0.697917 (0.014382) with: {'dropout_rate': 0.7, 'weight_constraint': 2} 0.687500 (0.008438) with: {'dropout_rate': 0.7, 'weight_constraint': 3} 0.704427 (0.011201) with: {'dropout_rate': 0.7, 'weight_constraint': 4} 0.696615 (0.016367) with: {'dropout_rate': 0.7, 'weight_constraint': 5} 0.680990 (0.025780) with: {'dropout_rate': 0.8, 'weight_constraint': 1} 0.699219 (0.019401) with: {'dropout_rate': 0.8, 'weight_constraint': 2} 0.701823 (0.015733) with: {'dropout_rate': 0.8, 'weight_constraint': 3} 0.684896 (0.023510) with: {'dropout_rate': 0.8, 'weight_constraint': 4} 0.696615 (0.017566) with: {'dropout_rate': 0.8, 'weight_constraint': 5} 0.653646 (0.034104) with: {'dropout_rate': 0.9, 'weight_constraint': 1} 0.677083 (0.012075) with: {'dropout_rate': 0.9, 'weight_constraint': 2} 0.679688 (0.013902) with: {'dropout_rate': 0.9, 'weight_constraint': 3} 0.669271 (0.017566) with: {'dropout_rate': 0.9, 'weight_constraint': 4} 0.669271 (0.012075) with: {'dropout_rate': 0.9, 'weight_constraint': 5} |

We can see that the dropout rate of 20% 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.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 |
# Use scikit-learn to grid search the number of neurons import numpy from sklearn.model_selection import GridSearchCV from keras.models import Sequential from keras.layers import Dense from keras.layers import Dropout from keras.wrappers.scikit_learn import KerasClassifier from keras.constraints import maxnorm # Function to create model, required for KerasClassifier def create_model(neurons=1): # create model model = Sequential() model.add(Dense(neurons, input_dim=8, kernel_initializer='uniform', activation='linear', kernel_constraint=maxnorm(4))) model.add(Dropout(0.2)) model.add(Dense(1, kernel_initializer='uniform', activation='sigmoid')) # Compile model model.compile(loss='binary_crossentropy', optimizer='adam', metrics=['accuracy']) return model # fix random seed for reproducibility seed = 7 numpy.random.seed(seed) # load dataset dataset = numpy.loadtxt("pima-indians-diabetes.csv", delimiter=",") # split into input (X) and output (Y) variables X = dataset[:,0:8] Y = dataset[:,8] # create model model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, verbose=0) # define the grid search parameters neurons = [1, 5, 10, 15, 20, 25, 30] param_grid = dict(neurons=neurons) grid = GridSearchCV(estimator=model, param_grid=param_grid, n_jobs=-1, cv=3) grid_result = grid.fit(X, Y) # summarize results print("Best: %f using %s" % (grid_result.best_score_, grid_result.best_params_)) means = grid_result.cv_results_['mean_test_score'] stds = grid_result.cv_results_['std_test_score'] params = grid_result.cv_results_['params'] for mean, stdev, param in zip(means, stds, params): print("%f (%f) with: %r" % (mean, stdev, param)) |

**Note**: Your results may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.

Running this example produces the following output.

1 2 3 4 5 6 7 8 |
Best: 0.714844 using {'neurons': 5} 0.700521 (0.011201) with: {'neurons': 1} 0.714844 (0.011049) with: {'neurons': 5} 0.712240 (0.017566) with: {'neurons': 10} 0.705729 (0.003683) with: {'neurons': 15} 0.696615 (0.020752) with: {'neurons': 20} 0.713542 (0.025976) with: {'neurons': 25} 0.705729 (0.008027) with: {'neurons': 30} |

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.

## Summary

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.

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

Thanks Yanbo.

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

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??

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

excellent post

Thanks Swapna.

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

It can if you set n_jobs=-1

Hi Jason,

In grid search, we do get train score right?

Why it’s not displaying in model.cv_results_ only test score we are getting..

You get a cross-validation score for each configuration tested.

Hi,

Great post,

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

Thanks!

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.

Hi Jason thanks for everything.

Could you explain what do you mean by repeatting structure in your reply please ?

Quick question on the GridSearchCV for CNN, param_grid=param_grid using the sklearn wrapper gives this error : ”ValueError: filters is not a legal parameter ”

How can we use the wrapper for the filters params of Conv1D ?

Thanks

Yes, see this post:

https://machinelearningmastery.com/review-of-architectural-innovations-for-convolutional-neural-networks-for-image-classification/

Perhaps try manually grid searching the parametres if you are working with time series, so that you can use walk forward validation:

https://machinelearningmastery.com/how-to-develop-deep-learning-models-for-univariate-time-series-forecasting/

Dear Jason,

This is an An excellent post. I have question: how can we grid search the optimum the number of filters in three different layers of CNN. For example: [60, 70 ,80] in layer 1, [20, 30, 40] in layer 2 and [5,10,20] in layer 3. I have searched everywhere for codes using grid search but could not find this. I really need to use grid search for this. I would be highly grateful for your kind advice. If possible, also reply in via my email address that I have provided (as this was a requirement for me to comment)

Thanks.

You might need to write some for-loops, e.g. do the search manually.

Also, we never find an “optimal” configuration, just a good enough configuration given the time/resources available.

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 grid.fit and model.fit 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.

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.

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

Thanks Vinay, I’m glad it’s useful.

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

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

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).

Hi, Jason

Thanks for your post. It is excellent.

I have a question.

You tune batch size and epoch first. But if you set a inappropriate number of neurons or activation function, then batch size and epoch tuning won’t make sense.

So I think we should tune all of these hyper-parameters at the same time.

How do you think about it?

They are all connected. If we could, we would tune all the parameters, but almost always it requires too many resources.

Hi Jason,

Do you recommend any particular order, which hyper parameter shoudl we tune first ?

There is an order in this article, you start with batch size and training epochs , then optimization etc.

Did you find any ressource or research paper, explaining the best consecutive tuning order ?

thanks

Learning rate!

Yes, I have many. Perhaps start here:

https://machinelearningmastery.com/learning-rate-for-deep-learning-neural-networks/

More here:

https://machinelearningmastery.com/start-here/#better

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 = grid.fit(x_train, y_train)

print(“Best: %f using %s” % (grid_result.best_score_, grid_result.best_params_))

Strange Satheesh, I have not seen that before.

Let me know if you figure it out.

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.

Glad to hear it.

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

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 = grid.fit(trX, 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.

Good question.

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

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?

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.

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:

json_file.write(model2json)

best_model.save_weights(best_model_file_path+’.h5′)

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

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.

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.

http://stackoverflow.com/questions/40572743/scikit-learn-grid-search-own-scoring-object-syntax

Thanks, Jan

Sorry Jan, I have not used a custom scoring function before.

Here are a list of built-in scoring functions:

http://scikit-learn.org/stable/modules/model_evaluation.html#scoring-parameter

Here is help on defining your own scoring function:

http://scikit-learn.org/stable/modules/model_evaluation.html#defining-your-scoring-strategy-from-metric-functions

Let me know how you go.

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

Excellent. Good luck Jan.

Good tutorial again Jason…keep on the good job!

Thanks Anthony.

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

http://machinelearningmastery.com/time-series-prediction-lstm-recurrent-neural-networks-python-keras/

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):

grid.fit(trainX, trainY)

grid.reset_states()

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

Thanks,

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.

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(Activation(‘relu’))

model.add(Dense(1, init=’normal’))

model.add(Activation(‘linear’))

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

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.

Has there been a blog post on this?

Not yet, maybe it’s time.

Have you solved this issue? I’m exploring Keras now as wel and came across exactly the same problem.

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.

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

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

Thanks

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.

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(neurons2))

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]

neurons2=[0,1,2]

param_grid = dict(neurons1=neurons1, neurons2=neurons2)

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

grid_result = grid.fit(X, Y)

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

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

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

Excellent, glad to hear it.

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.

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.

I’ll try to find the time.

Hi Jason,

Many thanks for this awesome tutorial !

I’m glad you found it useful Rajneesh.

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 = grid.fit(X, 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!

same issue here,

great tutorial, life saver.

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?

UPDATE:

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.

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!

It’s a bug in Keras 1.2.1. You can either downgrade to 1.2.0 or get the code from their github (where they already fixed it).

Yes, I have a write up of the problem and available fixes here:

http://stackoverflow.com/questions/41796618/python-keras-cross-val-score-error/41841066#41841066

Thank you so much for your help!

Jason,

Can you use early_stopping to decide n_epoch?

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

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.

Hi Jayant,

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

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？

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:

http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html

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

http://scikit-learn.org/stable/modules/model_evaluation.html#scoring-parameter

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

Thank you so much for your help!

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!

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.

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?

Thanks

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.

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

I’m here to help, if I can Dan.

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.)

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

Hi, Jason. I am getting

/usr/local/lib/python2.7/dist-packages/keras/wrappers/scikit_learn.py in check_params(self=, params={‘batch_size’: 10, ‘epochs’: 10})

80 legal_params += inspect.getargspec(fn)[0]

81 legal_params = set(legal_params)

82

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’

86

87 def get_params(self, _):

88 “””Gets parameters for this estimator.

89

ValueError: epochs is not a legal parameter

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

I experienced the same problem.I upgraded my keras and the same problem still occurs.

I was getting the ‘not a legal paramater’ error when I was trying to pass required inputs into my create_model function in the wrapper.

model = KerasClassifier(build_fn=create_model(input_dim = x ), verbose=0)

when I removed it and included it in the grid search instead it ran fine, I just added it to the dictionary of parameters

input_dim = [x]

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

Thanks Usman.

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

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

http://imgur.com/a/JJ7f1

The optmization of network topology ,learning rate ,batch size and epochs are done in stages?sir please tell me why these were done in stages

To make the explanation to the reader simpler.

Also probably to reduce search space, and thus computational time.

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?

Hi Daniel, yes great point.

Neural networks are stochastic and give different results when evaluated on the same data.

Ideally, each configuration would be evaluated using the average of multiple (30+) repeats.

This post might help:

http://machinelearningmastery.com/randomness-in-machine-learning/

Hi Jason!

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

Thanks

I have an example here:

http://machinelearningmastery.com/exploratory-configuration-multilayer-perceptron-network-time-series-forecasting/

Thank you so much Jason!

I’m glad it helped Pradanuari.

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\scikit_learn.py 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., …, 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))

139

–> 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

Regards

Ibrahim

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

Ok, I think your code needs to be placed after

if __name__ == ‘__main__’:

to work with multiprocess…

But thanks for the post is great…

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

n_jobs=-1 doesnt work on Windows.

@Ibrahim: Can you please explain, what part of the code needs to be behind

if __name__ == ‘__main__’: )

Assuming you have got several functions (i have a single python script acting as main file and the other stuff in a separate file, but at least functions like Jason does) you need to put this at the very begining of your main routine where everything comes together and is set-up. Note, since it is an if-condition, you need to tab everything below the condition.

@Jason maybe you can add this in the section where you talk about the problems on parallelization as a hint for windows users.

Thanks. I really don’t know about windows.

I’ve not seen a windows box in a long time and I’m impressed people use them for software development.

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?

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

Thanks a lot Jason!

Hello,

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.

(http://cs231n.github.io/assets/nn3/learningrates.jpeg)

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?

Thanks!

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.

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?

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

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

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…

Great question,

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

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.

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

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!

Please ignore my previous reply.

I find an answer here: https://github.com/fchollet/keras/issues/6451

Right now, the GridsearchCV using the scikit wrapper for network with multiple inputs is not available.

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

np.random.seed(2017)

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

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.

the command of autoencoder.compile is modified as the follows:

# Compile model

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

Can we do this for functional API as well ?

Perhaps, I have not done this.

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.

https://stackoverflow.com/questions/28005307/gridsearchcv-no-reporting-on-high-verbosity 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.

Thanks for the tip.

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?

Great question, see this paper:

https://arxiv.org/abs/1206.5533

Thanks Jason, I’ll check it out.

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 https://goo.gl/Q9Xy7B as a potential avenue using Spark (no Keras though).

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

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

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,

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:

http://machinelearningmastery.com/evaluate-skill-deep-learning-models/

Does that help?

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?

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.

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?

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.

Hi Jason,

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

GridSearch supports k-fold cross-validation by default. That is what the “CV” is in the name:

http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html

So sklearn has no GridSearch without cross validation?

In any case I found kind of a hack here to get rid of cv:

https://stackoverflow.com/questions/44636370/scikit-learn-gridsearchcv-without-cross-validation-unsupervised-learning

You can configure the k in CV to 1 to it does train/test. Then configure it to repeat.

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: https://www.dropbox.com/s/57n777j9w8bxf4t/keras_grid.py?dl=0

Any tip? Thank you

Hi Dr. Brownlee,

When I run this in Spyder IDE nothing happens after grid.fit.

It just appears to do nothing.

Any suggestions as to why?

Consider running from the command line.

The grid search may take a long time.

Hello Dr Brownlee,

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

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

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).

Consider running the example from the command line instead.

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

Yes.

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 model.fit 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…

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.

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

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.

Can i use this grid search without using keras model

For sure!

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.

Great suggestion, thanks.

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!

Very odd results. Double check your train/test data.

Also, see this post for a long list of ideas to try:

http://machinelearningmastery.com/improve-deep-learning-performance/

Thanks

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?

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.

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 :/

THanks.

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

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)

gsv_model.fit(x_train,y_train)

Maybe, you must have a test harness that you can trust, then explore different configurations of your model.

I have more on robustly evaluating neural nets here:

https://machinelearningmastery.com/evaluate-skill-deep-learning-models/

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?

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

{‘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?

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

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

Does that help?

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 ?

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?

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?

Yes, 1 run is one CV pass (k folds).

Jason,

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).

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

Jason, managed to get the inherited class working perfectly now:

class MLP_Classifier(KerasClassifier):

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

self.sk_params = sk_params

super().__init__(build_fn=None, **sk_params)

def __call__(self, callbacks=None, layer_sizes=None,activations=None,input_dim=0,init=’normal’,optimizer=’adam’, metrics=’accuracy’, loss=’binary_crossentropy’, use_dropout_input=False, use_dropout_hidden=False):

“””

Constructs, compiles and return a Keras model

Implements the “build_fn” function

Returns a “Sequential” model

“””

# Code to build a model (that would typically go in “build_fn”) goes here.

return model

Well done!

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.

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.

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%.

Thanks

P.S:

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 (cs231n.github.io/convolutional-networks/#case). 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

set_session(tf.Session(config=config))

Thanks!

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:

https://machinelearningmastery.com/evaluate-skill-deep-learning-models/

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.

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 ?

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’}

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

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?

Thanks

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.

I started looking at the pipeline (http://scikit-learn.org/stable/modules/pipeline.html) 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’}

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:

https://machinelearningmastery.com/automate-machine-learning-workflows-pipelines-python-scikit-learn/

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.

Thanks!

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 grid.fit([input1,input2],y)

Do you have any suggestion to handle it?

Thanks,

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

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?

What do you mean by valid labels exactly?

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.

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

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.

Hi Jason, Very nice tutorial..very well explained

Thanks.

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?

Wonderful question.

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

Hi Jason, by “number of repeats” do you mean to just repeat the process many times, with perhaps a different random seed?

Exactly.

Thank you for this great tutorial! I tried to adapt the code for a CNN, but I am running constantly in the same error. May anyone help?

That is the code:

def create_model(nb_filters=3, nb_conv=2, pool=20):

model = Sequential()

model.add(Convolution1D(nb_filters, nb_conv, activation=’relu’,

input_shape=(X.shape[1], X.shape[2]), padding=”same”))

model.add(MaxPooling1D(pool))

model.add(Flatten())

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

model.compile(loss=’binary_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])

model.summary()

return model

model = KerasClassifier(build_fn=create_model(), verbose=0)

nb_conv = [2, 4, 6, 8, 10]

pool= [10, 20, 30, 50]

param_grid = dict(nb_conv=nb_conv, epochs=pool)

grid = GridSearchCV(estimator=model, param_grid=param_grid)

grid_result = grid.fit(X, y)

And the error I am getting is “nb_conv is not a legal parameter”. Unfortunately, I do not understand why.

The API has changed:

https://keras.io/layers/convolutional/

Hi Jason,

Great post and Thank you.

What do you think is the best sequence when tuning all those Hyperparameters? I think difference sequence will lead to difference final Hyperparameters..

This post has some ideas (at the end):

https://machinelearningmastery.com/gentle-introduction-mini-batch-gradient-descent-configure-batch-size/

Also see the referenced paper.

Hi Jason,

What a great blog, I very much appreciate you sharing some of your expertise!

I want to grid search the hyperparams from my CNN, but I’m using data augmentation with ImageDataGenerator. So I’m not calling model.fit but model.fit_generator for the actual training.

This does not seem to be supported through the grid search..

Am I forced to write my own KerasClassifier implementation?

Would you advise to just fall back to using (nested) for loops instead, or would I be missing some ‘magic’ from the existing scikit gridsearch?

I would recommend writing your own for loops to grid search instead.

Hey Jason!

Needed help with model improvement!

Can you help me in understanding how to realize whether your model is suffering from

bad local minima, vanishing/exploding gradient problem?

If you have exploding or vanishing gradients, then you will have NaN outputs.

This post will give you ideas on how to lift skill:

http://machinelearningmastery.com/improve-deep-learning-performance/

This post will give you advice on how to effectively evaluate your model:

https://machinelearningmastery.com/evaluate-skill-deep-learning-models/

NaN outputs as in my predictions ?

Or the weights ?

If exploding gradient then weight will be very large (probably NaN) hence output would also be NaN.

But how will this logic be used for vanishing gradients. I this case the weights basically stop changing r8?

Should I use some kind of code that checks by how much the weights at each layer are changing…and if after a certain threshold they haven’t changed by a certain amount, I’ll declare vanishing gradient !

Try gradient clipping on the optimization algorithm.

I have a question for you, Jason and for general audience. I tried to find optimal number of neurons for one of the hidden layers. i did loop over my function which contains my deep learning model. It is fast enough for the values I define and I get a result based on accuracy. However, when I use your code, it is extremely slow and never reached to an end. How long does it take on your computer?

You could try to test fewer parameters or try to search on a smaller dataset?

Hey Jason,

Thank you for your quick reply. I try grid search for number of neurons on Iris data set for the purpose of learning. I scale the data first and then transform and encode the dependent variable. However, first of all, even though I use small data set or fewer parameters, it is slow; second of all, when I get the results, it is all zero. This is very basic example and I am pretty much sure that my code is correct but I guess I am missing out something.

Best: 0.000000 using {‘neurons’: 3}

0.000000 (0.000000) with: {‘neurons’: 3}

0.000000 (0.000000) with: {‘neurons’: 5}

THE CODE:

from pandas import read_csv

import numpy

from sklearn.preprocessing import LabelEncoder

from sklearn.preprocessing import StandardScaler

from keras.wrappers.scikit_learn import KerasClassifier

from keras.models import Sequential

from keras.layers import Dense

from keras.utils import np_utils

from sklearn.model_selection import GridSearchCV

dataframe=read_csv(“iris.csv”, header=None)

dataset=dataframe.values

X=dataset[:,0:4].astype(float)

Y=dataset[:,4]

seed=7

numpy.random.seed(seed)

#encode class values as integers

encoder = LabelEncoder()

encoder.fit(Y)

encoded_Y = encoder.transform(Y)

#one-hot encoding

dummy_y = np_utils.to_categorical(encoded_Y)

scaler = StandardScaler()

X = scaler.fit_transform(X)

def create_model(n_neurons):

model = Sequential()

model.add(Dense(n_neurons, input_dim=X.shape[1], activation=’relu’)) # hidden layer

model.add(Dense(3, activation=’softmax’)) # output layer

model.compile(loss=’categorical_crossentropy’, optimizer=’adam’, metrics=[‘accuracy’])

return model

model = KerasClassifier(build_fn=create_model, epochs=100, batch_size=10, initial_epoch=0, verbose=0)

# define the grid search parameters

neurons=[3, 5]

#this does 3-fold classification. One can change k.

param_grid = dict(n_neurons=neurons)

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

grid_result = grid.fit(X, dummy_y)

# summarize results

print(“Best: %f using %s” % (grid_result.best_score_, grid_result.best_params_))

means = grid_result.cv_results_[‘mean_test_score’]

stds = grid_result.cv_results_[‘std_test_score’]

params = grid_result.cv_results_[‘params’]

for mean, stdev, param in zip(means, stds, params):

print(“%f (%f) with: %r” % (mean, stdev, param))

Sorry, I cannot debug your code/problem for you.

I totally understand you. Thank you so much, though. I figured out my mistake. Iris dataset is very well balanced so I need to shuffle the data because GridSearchCV is using 3-Fold Cross Validation.

Glad to hear it.

Thanks for sharing such a wonderful tutorial. Learnt many new things.

How can i save all the models that the grid search is generation with identifiers for each model?

I am an R user. This how I do it in R to save models with passing parameter values to its names.

xgb.object <- paste0('/path/xgb_disc20_new_',

sample.sizes[i], '_', s,'_',nrounds[j],'_',max.depth[k],'_',eta[l], '.RData')

write.table(cbind(sample.sizes[i], s,nrounds[j],max.depth[k],eta[l],tpr, tnr, acc, roc.area,

concordance), paste0('/path/xgb_disc20_new_', min.sample.size,'_', max.sample.size,

'.csv'), append=TRUE, sep=",",row.names=FALSE,col.names=FALSE)

How can this be achieved in python for keras(neural network) and other models in other libraries?

I would recommend using grid search to find the parameters for a well performing model then train a new standalone model with those parameters that you can then save.

thank you jason for your quick reply . I will try that way.

Hi Jason,

Thank you for the great tutorial. I just have an issue when using exactly your code: when I try to parallelize the grid search with n_jobs=-1, I end up with the error “AttributeError: Can’t get attribute ‘create_model’ on ” while it works well without parallelization. Any idea where the issue comes from?

Thank you,

Wassim

I’m not sure, perhaps you cannot parallelize the grid search with Keras models.

Hi Jason,

The example code calculates the best score for accuracy to obtain the hyperparameter.

In my problem, I want to find RMSE rather than accuracy because it is regression problem (numerical prediction).

However, ‘grid_result.cv_resluts_’ only provides ‘fit_time’ and ‘score’, so it can not calculate RMSE.

What should I do?

Thank you.

You can change the configuration to calculate MSE (e.g. scoring=’neg_mean_squared_error’) and then take the square root.

Learn more here:

http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html

Hi Jason,

Thank you for this post.

Is there anything that prevents me to use Grid Search with train_on_batch() instead of fit()?

Thank you for letting me know.

All the best,

Estelle

I think the wrapper is quite limited and does not offer this facility via sklearn.

Thanks for your quick answer.

All the best,

Estelle

No problem.

Thanks very much for the tutorial. It is extremely helpful for my work. I came across a problem with grid search with Keras (tensorflow backend). I want to run the same grid search on different datasets. Everything works fine on the first dataset. But when I fit the grid search to the second dataset, the program got stuck there. I run the grid search with n_jobs=-1 and put keras.backend.clear_session() between two fits. You can replicate this issue by fit to the data twice in your examples. Could you please kindly help me with this issue?

I’m sorry to hear that, perhaps change n_jobs to 1?

Thanks for the quick reply. It works when n_jobs=1, but I do need parallel threads for speed.

The neural network will be using all the cores, so running multiple threads may not offer any benefit.

I got it to work by just fitting one dataset in the python script and looping the python script over multiple datasets in a bash script. I am still not clear why second fitting fails in python, but this is a not-so-beautiful workaround.

Glad to hear that you made some progress.

Hi Jason

Thank you so much for sharing your knowledge.

I am trying to optimize the number of hidden layers.

I can´t figure it out how to do it with keras (actually I am wondering how to set up the function create_model in order to maximize the number of hidden layers)

Could you please help me?

Thank you

Perhaps the number of layers could be a parameter to your function.

Hi Jason,

Thanks for this insightful and useful tutorial as always

No doubt your blog posts are arguably the best in the field of data sciences

Best wishes

Thanks Sean.

Hello Jason,

I decided to try the code on a textual data of about 3000 tweets having binary classification (Y) and the text corpus as (X). Started off with tuning the batch size and number of epochs

but got the following error:

Here’s the modified code below:

Thanks

Sorry to hear that, it’s not clear to me. Perhaps post to stackoverflow to get help debugging your code?

Hi Jason, first thanks for your articles! Super useful!

I tried to execute the gripsearch but cam up with parallelism issues. I have a Windows OS and I get this error when I try to run the script on multiple cpus:

ImportError: [joblib] Attempting to do parallel computing without protecting your import on a system that does not support forking. To use parallel-computing in a script, you must protect your main loop using “if __name__ == ‘__main__'”. Please see the joblib documentation on Parallel for more information.

Do you know how I should address that?

Thanks in advance

Perhaps try setting the number of jobs to 1?

Hi Jason! Yes this works but it is very slow as this is not parallel. Do you understand why it cannot run in parallel and how to fix that?

Thanks again !

Olivier

The backend is parallelized and the two levels of parallelization are in conflict.

Thanks a lot for such a wonderful post. Overall, there are a lot of parameters that need to be tuned. I was thinking to use RandomizedSearchCV instead of GridSearchCV. Still, it will be time consuming for a lot of simulations. Do you have any suggestion for fast parameter tuning? For example, can we say that specific parameters have more effect on scores, so lets try to Grid/RandomizedSearchCV them first?

Yes, there are some great tips at the end of this post:

https://machinelearningmastery.com/gentle-introduction-mini-batch-gradient-descent-configure-batch-size/

Dear Jason,

Fantastic post, thank you for this wonderful tutorial.

I was wondering if it would be more appropriate to tune all the hyperparameters at one go instead of breaking it up into various parts as shown above – you may be doing it for the sake of visibility of how each component is tuned but would it be better to tune everything together since there might be “interactions between the hyperparameters” which would not be captured if they were tuned separately?

If you have the resources, then sure.

Hi Jason,

Many thanks for a series of excellent posts!

I have an extremely imbalanced data set to study, of which #negative : #positive is about 100:1. When I built the first model, I performed 10-fold validation and in each validation round, I use oversampling to add positive samples on training data, but not on testing data. Now I question is: if I want to perform hyperparameter search, how do I tell GridSearchCV() to do oversampling for each round of cross-validation?

Many thanks

Good question, you might need to use a Pipeline and have data prep happen within it.

Hello Jason

A good 2018 to you. I have a question about how Keras early stopping callbacks might be able to use the GridSearchCV k-fold generated validation data set as their

`val_loss`

or`val_acc`

. The question I posted on StackOverflow but I wished to call your attention to it – should you so wish.https://stackoverflow.com/questions/48127550/how-do-i-implement-early-stopping-with-keras-and-the-sklearn-gridsearchcv-cross

Kind regards,

Justin

I would suggest not combing CV and early stopping.

Could early stopping be used as a substitute for grid searching epoch size?

Yes, but you might need to code it up yourself. sklearn might blow up.

Hello sir

if i have large dataset the also we can do this hyperparameter tunning .

If i have 70 to 80 feature column and about 50000 rows.

can we apply this tunnig

Sure, you might need a large computer or to split the work up across many computers.

Perhaps you can work with a sample of your data.

how to select the hidden layer if i have largedataset mentioned as above

I don’t follow, what do you mean exactly?

Very good post.

Hyper Parameter Tuning: How can I do grid search on number of neuron/epochs or batch size using Keras interface in R.

Sorry, I don’t have an example in R.

Hi,I am facing a basic query where i have training and test set.i built lstm on training and using history = model.fit(trainX, trainY, epochs=100, batch_size=50,

validation_data=(testX, testY), verbose=0, shuffle=False) to fit my model.

After this i tried to model.predict(testX) to get predicted Y values.Now that was basic code.i am now trying to apply gridsearch.what variation in the history statement code i have to make to apply grid =

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

grid_result = grid.fit(testX, testY, verbose=0, shuffle=False)

can gridsearchcv work for time series as well?

Not really. You will have to write your own for loops and perform walk forward validation.

Hi Jason, thank you for your great tutorial! My question here is about ‘grid_result.best_score’. In this article the best score seems to be the best mean score, but in a regression problem, the mean score is irrelevant, so I have to look for the best std score. Is that correct?

Mean score in regression will be mean error. Not irrelevant.

I see. But when I run the code, the ‘grid_result.best_score’ printed out the biggest score. I don’t think that’s right, cause in a regression problem I should look for the smallest mean error. Am I understanding this right?

Below are the results:

Best: 0.062234 using {‘optimizer’: ‘Nadam’}

0.059561 (0.017101) with: {‘optimizer’: ‘SGD’}

0.056818 (0.013662) with: {‘optimizer’: ‘RMSprop’}

0.059617 (0.014734) with: {‘optimizer’: ‘Adagrad’}

0.061506 (0.014503) with: {‘optimizer’: ‘Adadelta’}

0.059331 (0.014835) with: {‘optimizer’: ‘Adam’}

0.057696 (0.014828) with: {‘optimizer’: ‘Adamax’}

0.062234 (0.010834) with: {‘optimizer’: ‘Nadam’}

Yes, for regression it should be the smallest error. Are you using negative mse as the score function?

I’m not sure. I just copied the codes from this tutorial and changed ‘KerasClassifier’ to ‘KerasRegressor’.I didn’t make any change other than that. I don’t understand how score function works and I’m not familiar with the concept of negative mse. Would you please elaborate?

You must specify a scoring function in sklearn, learn more about the API here:

http://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html

Here are examples:

http://scikit-learn.org/stable/modules/model_evaluation.html#scoring-parameter

very good tutorial, But I have a small question. can I tune all these hyperparameters together or I should take a part of the dataset and tune them separately like the examples you mentioned.

Ideally, you would tune them all together, but this is often to computationally expensive.

Is there a way to do similar things in R using the Caret package? Or other package that can help you with hyperparameter grid search when using Keras in R?

I don’t know if Keras and caret are compatible, sorry.

hi Jason,

do i need to split the training data for cross validation, or only perform splitting on the input data.

Why do you want to split exactly? You goals will help me answer your question.

Thanks Jason for the quick reply…i will figure that out.. Just another minor question, is there any way to perform data preprocessing on 3d input (due to the input shape for lstm)

Sure, but it might be easier (or make more sense) to perform data prep prior to shaping data for the LSTM.

Thanks Jason..i will try that out.. Is it a good idea to tune the hyperparameter using the keras wrapper, then apply those tuned parameters on lstm model? Hope to get some comments on it. Thank you.

You can. Or you can write your own for loop and tune the model directly.

Thanks a lot Jason.. i will definitely try that one out..

Hi Jason, wonderful post. I love your books – amazing.

I wish to include callbacks in the Grid Search (one for TensorBoard and one for logging losses on every combination over the params).

I have something like:

loggerCB = keras.callbacks.TensorBoard(log_dir=’logs’, histogram_freq=0, write_graph=True)

class LossHistory(keras.callbacks.Callback):

def on_train_begin(self, logs={}):

self.losses = []

def on_batch_end(self, batch, logs={}):

self.losses.append(logs.get(‘loss’))

historyCB = LossHistory()

grid_search = GridSearchCV(estimator=model,

param_grid=fit_params,

scoring=’accuracy’,

cv=10)

grid_search = grid_search.fit(X_train, y_train, fit_params={‘callbacks’: [loggerCB, historyCB]})

BUT I got this error:

TypeError: Unrecognized keyword arguments: {‘fit_params’: {‘callbacks’: [, ]}}

How can I pass callbacks using Grid Search?

Thanks,

Boris Branson

ry, I have not used callbacks with a grid search. You might need to write your own for-loops for the search.

Hello Jason,

let me congratulate for the good post.

I am curious about the use of CV . Each time you call

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

you are compiling a new keras model with the new set of parameters.

Are these different models of keras, compiled one after another, accumulating in the memory? Would this imply a memory usage problem in the case of an extensive grid search with bigger models? Any tips?

Best,

Alessandro

Yes, each model is evaluated and discarded.

For larger models, you could run each fold on a different machine (e.g. run the eval manually).

Hello Jason,

I see you have used only SGD in the example of learning rate parameterization. Is it possible to combine different values for lr with different optimizers (not only SGD) in one grid search or i’d need a for loop?

Yes, but the more parameters you grid search at once, the slower the search.

I Jason your article is super useful, but I am having problem using it for MNIST data set which is a three dimensional data set , When I try so ‘fit’ this one gives me error, Dimension error. Can you do one for MNIST data set. Thanks a lot

Perhaps try this tutorial:

https://machinelearningmastery.com/handwritten-digit-recognition-using-convolutional-neural-networks-python-keras/

Hi Jason, Great tutorial, always learn a lot from your post. I have question, is it possible to combine all the parameters and with gridsearch? Seems more than thousands of combinations. For some models it will cost few days or weeks. Is there any better solution for this? randomgridsearch or something else? Thanks again!

Yes, but as you say, you will need a lot of time or a lot of parallel compute resources to get a result.

Random search is often preferred because you can uniformly sample the domain and get good enough results quickly.

Thanks for your reply. Googled a lot but didn’t find any method to search optimizers and their params, say different optimizer, adam and it’s learning rates. Is there any suggestions? Thanks!

Yes, just start searching for viable params on your model/data. No need to find confirmation.

Hello Jason,

Thanks for this awesome tutorial. Am very fresh in machine learning and your tutorials are so simplified and easy to follow.

Am encountering an error when i run the epochs and batch size tuning code. Kindly help

This the code part bringing the error…

# create model

model = KerasClassifier(build_fn=create_model, verbose=0)

# define the grid search parameters

batch_size = [10, 20, 40, 60, 80, 100]

epochs = [10, 50, 100]

param_grid = dict(batch_size=batch_size, epochs=epochs)

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

grid_result = grid.fit(X_train, y_train)

TypeError: __call__() missing 1 required positional argument: ‘inputs’

Sorry, I have not seen this error. Are you able to confirm that you have copied all of the code and that your development environment is up to date?

I ran into this over the weekend, and hopefully to same some one else some pain down the road:

I kept getting the following error when working the prediction section of my code, which frankly was driving me nuts:

TypeError: call() missing 1 required positional argument: ‘inputs’

After researching the error message I came upon this comment which let me to the resolution:

_The thing here is that KerasRegressor expects a callable that builds a model, rather than the model itself. By wrapping your function in this way you can return the build function (without calling it)._ [Source](https://stackoverflow.com/questions/47944463/specify-input-argument-with-kerasregressor)

Solution: I needed to **wrap** my

`buildModel()`

function! 🙁Once I ‘wrapped’ the

`buildModel()`

function the prediction code blocks finally started working. Git it a try, and it should resolve your issue. The link I provided above should give you a working code example. If not let me know, and I’ll post my working example for you.Thanks!

It might be easier to write your own for loops to grid search Keras models.

dear jason

how much time this program run while tunning ?like tuning epoch and batch size?

It depends on the size of the dataset, the size of the model and the speed of your system.

Hi,

As you mention in your blog “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.” does this mean we should so the hyper parameter search in one grid instead of dividing.

regrads,

Yumlembam Rahul

Ideally, if you have the time and resources.

sir,

I have tried above code. it is executing ,but not displaying results..i don’t know the reason ..

Perhaps try from the command line, then be patient.

Perhaps try to reduce the data set size or use fewer combinations?

hi, in your example optimizer parameter are not specified while doing grid search do they assume default values if not specified??

and for reproducibility of result i added the following code and have been able to get same result

import os

os.environ[‘PYTHONHASHSEED’] = ‘0’

np.random.seed(42)

rn.seed(12345)

session_conf = tf.ConfigProto(intra_op_parallelism_threads=1, inter_op_parallelism_threads=1)

from keras import backend as K

# The below tf.set_random_seed() will make random number generation

# in the TensorFlow backend have a well-defined initial state.

# For further details, see: https://www.tensorflow.org/api_docs/python/tf/set_random_seed

tf.set_random_seed(1234)

sess = tf.Session(graph=tf.get_default_graph(), config=session_conf)

K.set_session(sess)

sir,

I have doubt ,Whether LSTM concept could be used for prediction of diabetes dataset(PIMA INDIAN DATASET)…I don’t know how LSTM Learns data from dataset..is it possible to put an hands on calculation..

LSTMs are not appropriate for classification problems. They are intended for sequence classification:

https://machinelearningmastery.com/sequence-prediction/

Is it possible to put an hands on calculation particularly for hidden layers and LSTM layers..Is it possible to put manual calculation on weights(how it transfer weight from one layer to another layer)…

Sure, but you will need to code these as extensions to the Keras library.

sir ,

i have tried above code without n_jobs==-1 parameter .it is working …I have doubt ,that is above code can be run using LSTM model …is that possible…

Perhaps set it to 1 thread and let Keras have all of the cores?

Hi Jason,

I’m sure it’s possible – but I can’t figure it out.

The above code gives me as a result the best hyper-parameters as measured on the cross-validation.

Now which adjustments to the code would be necessary to additionally calculate the optimum hyper-parameters on a test set?

The optimum hyper-parameters seem to lead to significantly different results when applied to my model that I use to predict values.

Thanks

Max

sir ,

I have an doubt that is multivariate time series data can be used for classification or prediction .whether we can use that data for prediction or classification or both.

You can learn the difference between classification and regression here:

https://machinelearningmastery.com/classification-versus-regression-in-machine-learning/

sir,

In LSTM model you are using only RMSE loss function …..why you are not used other loss function ..In particular sequence prediction problem (forecasting) you used only RMSE loss function ….why sir.

I use MSE not RMSE. You can try other loss functions if you prefer. I find MSE loss function works well for most problems.

Hi

Thanks for your nice post.

Could you please let me know how to incorporate class_weight and tune it?

Sorry, I do not have a worked example.

Hello, great post as always!

I had a query regarding this. So I have a training set and a test set, and I am using a stacking ensemble for predictions.

So when I run GridSearchCV on this, should I fit just the training set on this and print CV score on the training set ONLY? And not touch the test set at all?

Also should I fit the new grid classifier on the set before printing the CV score or after?

Yes, hold back the test set, and use the training set for CV.

More on this here:

https://machinelearningmastery.com/difference-test-validation-datasets/

model = KerasClassifier(build_fn=create_model, verbose=0)

# define the grid search parameters

batch_size = [10, 20]

epochs = [10, 20, 30]

param_grid = dict(batch_size=batch_size, epochs=epochs)

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

grid_result = grid.fit(x_train, y_train)

When I am running this code snippet I am getting error as

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

Can you please help me on that ?

Sorry, I have not seen this error before.

Hi Jason,

First and foremost, this is an incredible writeup – very informative.

I’m getting an error that reads “can’t pickle _thread.RLock objects”

When I use the following code:

————————————————————————–

def build_neural_network(n_predictors, hidden_layer_neurons):

“””

Builds a Multi-Layer-Perceptron utilizing Keras.

Parameters:

x_train: (2D numpy array) A n x p matrix, with n observations

and p features

y_train: (1D numpy array) A numpy array of length n with the

target training values.

hidden_layer_neurons: (list) List of ints for the number of

neurons in each hidden layer.

Returns:

model: A MLP with 2 hidden layers

“””

model = Sequential()

input_layer_neurons = n_predictors

model.add(Dense(units=hidden_layer_neurons[0],

input_dim=input_layer_neurons,

kernel_initializer=’uniform’,

activation=’relu’))

model.add(Dense(units=hidden_layer_neurons[1],

kernel_initializer=’uniform’,

activation=’relu’))

model.add(Dense(units=1))

model.compile(optimizer=’rmsprop’,

loss=’mse’)

return model

# columns variable defined elsewhere, works as expected

mlp = build_neural_network(len(columns), [8, 12])

model = KerasRegressor(build_fn=mlp)

# create parameter lists for GridSearchCV

batch_size = list(np.arange(10, 250, 10))

epochs = list(np.arange(5, 20, 5))

neural_net_grid_dict = {‘batch_size’: batch_size,

‘epochs’: epochs}

neural_net_grid = GridSearchCV(estimator=model,

param_grid=neural_net_grid_dict,

scoring=’neg_mean_squared_error’,

verbose=1,

n_jobs=-1)

mask = df[‘Date’] == ‘2006-11-06’

X, y = create_X_y(df[mask], columns)

grid_result = neural_net_grid.fit(X, y)

——————————————————–

Any idea what might be going on?

Sorry, I have not seen this error. Perhaps try posting to stackoverflow?

Thanks so much ! This post helped me a lot !

I’m glad to hear that.

I am experiencing the same error “can’t pickle _thread.RLock objects”, may I know how you solved it?

Hi Jason,

how can tune your model to found hyperparameters (learning rate, epoch and output dim in hidden layer) using RandomizedSearchCV?

Thanks !!

Regards

Juan

Specify ranges and search. What is the problem exactly?

Hi Jason, I got a help from this blog post. Thank you very much!

I have one question though. What if I want to test with optimizers that has customized parameters and not default parameters. From your example, it’s just an array of Strings of optimizers name.

Do you know how I can do this?

Best,

June

You can provide lists of strings with optimizer names if you wish.

Yes. Isn’t this what’s provided in the example code?

optimizer = [‘SGD’, ‘RMSprop’, ‘Adagrad’, ‘Adadelta’, ‘Adam’, ‘Adamax’, ‘Nadam’]

What I meant was not with default ones but like when I have my own optimizer defined as follows:

sgd_custom = SGD(lr_rate=0.7)

adam_custom = (decay=0.005)

How can I give optimizer list for this setting? optimizer=[sgd_custom, adam_custom]?

Good question.

Yes, you could provide a list of pre-configured objects to use instead of strings.

Hi Jason,

Your posts are really helpful – thanks a lot!

1. I’m using grid search on my own Keras CNN and everything is working. One thing that keep’s confusing me though: The F1 measures reported by grid search are always a bit (3-4%) higher than when running the same network configurations in Keras directly. I know that Keras isn’t using CV, but this shouldn’t lead to systematic deviations in one direction but to deviations in both directions I think.

2. Also I found that my network is always performing slightly better (accuracy) when using the TF-Layers API instead of Keras, even though the network configurations are exactly the same (as far as I can control this in Keras).

Any ideas why Keras seems to perform poorer? Have others experienced the same issues with Keras? I just can’t figure it out…

Cheers,

Philipp

No good idea sorry. It might be statistical chance, or it might be real. See if you can tease this out with some hypothesis tests on the results.

Thanks, Jason.

Just to let you know: Apparently it has something to do with the F1 score. Accuracy scores reported by grid search are pretty much the same as my results in Keras.

Interesting.

Hi Jason, thank you for very detailed and interesting tutorial.

1. I tried to grid hyperparameters of epochs and batch size as your code. No result was launched and no error message appeared. after that, i changed n_jobs equal 1, python gave me the result. I do not understand why value of n_jobs = -1 prevented the calculation process.

2. If i have complicated network (with two layers for example), could you tell me how grid can be implemented with number of epochs and batch size?

Thank you a lot!

Might have caused a deadlock internally.

I don’t understand your second question sorry, perhaps you can rephrase it?

Hi, Jason, excellent post and help lot for improving my predictive model.

I have one question, is there any way I can optimise number of layer in network ?

Yes, use a grid search and choose the configuration with the lowest loss.

I tried the gird search but got this error

ipython-input-49-ea7e264ec276> in ()

3 param_grid = dict(batch_size=batch_size, epochs=epochs)

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

—-> 5 grid_result = grid.fit(xs, testY)

6 # summarize results

7 print(“Best: %f using %s” % (grid_result.best_score_, grid_result.best_params_))

~\Anaconda3\envs\tfdeeplearning\lib\site-packages\sklearn\model_selection\_search.py in fit(self, X, y, groups, **fit_params)

612 refit_metric = ‘score’

613

–> 614 X, y, groups = indexable(X, y, groups)

615 n_splits = cv.get_n_splits(X, y, groups)

616 # Regenerate parameter iterable for each fit

~\Anaconda3\envs\tfdeeplearning\lib\site-packages\sklearn\utils\validation.py in indexable(*iterables)

196 else:

197 result.append(np.array(X))

–> 198 check_consistent_length(*result)

199 return result

200

~\Anaconda3\envs\tfdeeplearning\lib\site-packages\sklearn\utils\validation.py in check_consistent_length(*arrays)

171 if len(uniques) > 1:

172 raise ValueError(“Found input variables with inconsistent numbers of”

–> 173 ” samples: %r” % [int(l) for l in lengths])

174

175

ValueError: Found input variables with inconsistent numbers of samples: [17, 1]

I have some suggestions here John:

https://machinelearningmastery.com/faq/single-faq/why-does-the-code-in-the-tutorial-not-work-for-me

Hey,

what refers 8 in the input dim ? i have a time serie problem a dataset with 41 observation how could i deal with this ?

It refers to 8 input variables.

You could define a window of lag obs as input features. Perhaps experiment with different window sizes.

could we use only one hiden layer that contain lstm bloc. i want to grid search hyperparametre for my lstm achitecture how could i specify this on code.

Yes, you could adapt the above examples to search layers/nodes in an LSTM.

Astounding post, thank you! I wonder how I could evaluate the loss and accuracy evolution of the KerasClassifier according to epoch. Is there something like the history class returned from the model.fit method from SciKitLearn?

Not that I am aware, I believe you would need to use the Keras API directly and collect history objects from each run.

Dear Jason,

I found this article as very useful for my research. Thank you very much.

Is it possible to find the best CNN architecture (No.of layers, Kernel size, Kernel initialization, Pooling Technique etc) for a given dataset by using GridSearch or RandomSearch?

There is no “best”, just good enough based on the time and resources we have available.

Hello Jason Sir , i want to know that how could i apply CNN concept for non image data which contains large datasets in form of rows and coloumns , and how could i apply padding in 50,000 Rows and 20 coloumns , Kindly suggest an approach.

CNN is not appropriate unless there is some spatial relationship between the observations, e.g. time or space.

Hi

thanks for this post and the replies to questions.

I have a question on the properties of the cnn, if you have a dataset like the pollution dataset.

If we have one binary variable as target in a classification with 10 exogenous variables and it is a daily forecast.

Let us say we have 500 days of data.

I can create a multivariate timeseries forecast and have 5 timesteps in my window so that my train shape will be (500,5,10)

If I apply Conv1D, it should extract features out of all the 10 variables right ?

or does it apply a Conv1D on each exogenous variable separately.

What I try to understand is : does it capture interactions of exogenous variables ?

Does the Conv2D only work for images or for times series too ?

For each window of 5 timesteps, we have 5 timesteps and 10 exogenous variables so we could think this is 2D.

Thanks J

Yes, you van get stated here:

https://machinelearningmastery.com/how-to-develop-convolutional-neural-network-models-for-time-series-forecasting/

Hi

I think you are pointing me again to the same tutorial but my questions come from this one.

Questions see above.

Question 1 :

If I apply Conv1D, it should extract features out of all the 10 variables right ?

or does it apply a Conv1D on each exogenous variable separately.

Question 2 : does it capture interactions of exogenous variables ?

Question 3 :

Does the Conv2D only work for images or for times series too ?