Hyperparameters

From Wikipedia: “In machine learning, a hyperparameter is a parameter whose value is used to control the learning process. By contrast, the values of other parameters are derived via training.”

Setting Hyperparameters

In julearn, we can set the hyperparameters using the model_params parameter in run_cross_validation().

In order to specify a hyperparameter for a specific step of the pipeline, we need to prefix the parameter name with the step name, separated by double underscores (__). For example, if we want to specify the parameter kernel='linear' of the svm model, we need to do the following:

model_params = {'svm__kernel': 'linear'}
run_cross_validation(X, y, data=df, model='svm', model_params=model_params)

Another common example, setting the parameters for a PCA decomposition:

model_params = {'pca__n_components': 0.99}
run_cross_validation(X, y, data=df, preprocess_X=['zscore', 'pca'],
                     model='svm', model_params=model_params)

TODO: Add doc for confound and y transformers hyperparameters

Tuning Hyperparameters

Another useful technique is called hyperparameter tuning. This technique is normally applied when we want to set the hyperparameters in a data driven way.

In this case, we use yet another cross-validation scheme to split the training data and evaluate the best set of hyperparameters for our model.

In julearn, hyperparameter tuning is enabled by giving more than one option for at least one hyperparameter.

For example, obtaining the best regularization parameter (C) for the svm model can be done by testing several candidate values:

model_params = {
    'svm__C': [0.001, 0.01, 0.1, 1, 10, 100],
    'svm__kernel': 'linear'}
run_cross_validation(X, y, data=df, model='svm', model_params=model_params)

In this case, since there are more than option for C, julearn will evaluate which is the best value. For the kernel parameter, there will be no search as 'linear' is the only value specified.

There are three more parameters that we can use to specify how the hyperparameter search will be done:

• search: The kind of algorithm to apply. The value 'grid' (default) will apply a GridSearchCV. The value 'random' will use a RandomizedSearchCV

• cv: The cross-validation scheme to use for the hyperparameter tuning. The default is to use the same scheme as for the model evaluation.

• scoring: The scoring metric to optimize.

• search_params: Additional parameters for the search algorithm.

The following examples performs a randomized search with 6 options for C and 13 options for gamma. Testing all the combinations will require 78 different evaluations. The randomised RandomizedSearchCV will not try them all, but a number specified in n_iter. In this case, we want to try 50 random combinations.

model_params = {
    'svm__C': [0.001, 0.01, 0.1, 1, 10, 100],
    'svm__gamma': [1e-9, 1e-8, 1e-7, 1e-6, 1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1,
                   10, 100, 1000],
    'svm__kernel': 'rbf',
    'search': 'random',
    'search_params': {'n_iter': 50}

}
run_cross_validation(X, y, data=df, model='svm', model_params=model_params)

For more information, see scikit-learn’s Hyperparameter Tuning.