Changelog¶

v1.2.0dev¶

n_folds and n_iter parameters have been renamed to n_splits to conform to the sklearn API.

Added support for TimeSeriesSplit and LeavePOut cross-validators.
Improved osprey dump JSON output. The hyperparameters for each run are now stored along all the other settings in the same dictionary, allowing for subsequent easier loading and plotting.
Added max_param_suggestion_retries entry to the config file. This limits the number of times that strategy.suggest is called when attempting to produce a trial with a set of params not previously tested in the history.
Added n_jobs flag for osprey worker to control how many threads are used for cross-validation.
Added the ability to specify three different acquisition functions for the gaussian processes strategy: expected improvement ei, upper confidence bound, ucb and the original Osprey function (the default), osprey.
Refactored code: GP strategy has been replaced by a general Bayes strategy. This will eventually allow different surrogate functions, other than Gaussian Process, to be specified.

Fixed issue that was causing crashes when there was an attempt to write estimator parameters (e.g. numpy arrays) which couldn’t be serialized by JSON.
Fixed crashes when using jump variables of type int.
Fixed error in the way integer variables were selected from results of Gaussian processes search strategy.
sklearn.grid_search module has been replaced by sklearn.model_selection to conform to new sklearn API

Implemented Config.trial_results, allowing convenient retrieval of trials as a pandas.DataFrame (#190)

Fixed issue where random_seed was not passed to estimator (#198)
Fixed bokeh.io.vplot deprecation warning (#192)
Fixed ungraceful failures when using GP with a single choice in search space (#191)
Fixed parsing of jumpvar (#164)

This is the first stable version of Osprey