dival.hyper_param_search module

Optimization of hyper parameters.

Both grid search and random search using the hyperopt library are supported.

The hyper parameter specification of a reconstructor class, optionally including default options for optimization, are specified in the class attribute HYPER_PARAMS.

dival.hyper_param_search.optimize_hyper_params(reconstructor, validation_data, measure, dataset=None, HYPER_PARAMS_override=None, hyperopt_max_evals=1000, hyperopt_max_evals_retrain=1000, hyperopt_rstate=None, show_progressbar=True, tqdm_file=None)

Optimize hyper parameters of a reconstructor.

Parameters:

• reconstructor (Reconstructor) – The reconstructor.

• validation_data (DataPairs) – The test data on which the performance is measured.

• measure (Measure or str) – The measure to use as the objective. The sign is chosen automatically depending on the measures measure_type.

• dataset (Dataset, optional) – The dataset used for training reconstructor if it is a LearnedReconstructor.

• HYPER_PARAMS_override (dict, optional) – Hyper parameter specification overriding the defaults in type(reconstructor).HYPER_PARAMS. The structure of this dict is the same as the structure of Reconstructor.HYPER_PARAMS, except that all fields are optional. Here, each value of a dict for one parameter is treated as an entity, i.e. specifying the dict HYPER_PARAMS[...]['grid_search_options'] overrides the whole dict, not only the specified keys in it.

• hyperopt_max_evals (int, optional) – Number of evaluations for different combinations of the parameters that are optimized by hyperopt and that do not require retraining. Should be chosen depending on the complexity of dependence and the number of such parameters.

• hyperopt_max_evals_retrain (int, optional) – Number of evaluations for different combinations of the parameters that are optimized by hyperopt and that require retraining. Should be chosen depending on the complexity of dependence and the number of such parameters.

• hyperopt_rstate (np.random.RandomState, optional) – Random state for the random searches performed by hyperopt.

• show_progressbar (bool, optional) – Whether to show a progress bar for the optimization. Default: True.

• tqdm_file (file-like object) – File/stream to pass to tqdm.