v4.0.0

v4.0.0

v3.2.4

Changes from v3.0.0 -> v3.2.4:

• Decouple number of runs from active processes (Thanks to PartiallyTyped). This reduces memory load if number of jobs is huge
• New feature: The progress board enables the user to monitor the optimization progress during the run.
  • Display trend of best score
  • Plot parameters and score in parallel coordinates
  • Generate filter file to define an upper and/or lower bound for all parameters and the score in the parallel coordinate plot
  • List parameters of 5 best scores
• add Python 3.8 to tests
• add warnings of search space values does not contain lists
• improve stability of result-methods
• add tests for hyperactive-memory + search spaces

v2.3.0

• add Tree-structured optimization algorithm (idea from Hyperopt)
• add Decision-tree optimization algorithm (idea from sklearn)
• enable new optimization parameters for bayes-opt:
  • max_sample_size: maximum number of samples for the gaussian-process-reg to train on. Sampling done by random choice.
  • skip_retrain: skips the retraining of the gaussian-process-reg sometimes during the optimization run. Basically returns multiple predictions for next output (which should be apart from another)

v2.1.0

• first stable implementation of "long-term-memory" to save/load search positions/parameter and results.
• enable warm start of sequence based optimizers (bayesian opt, ...) with results from "long-term-memory"
• enable the usage of other gaussian-process-regressors than from sklearn. GPR-class (from gpy, GPflow, ...) can be passed to "optimizer"-kwarg

v2.0.0

API-change to improve usage. Class accepts training data. "search"-method accepts search_config and other optimization-run specific arguments like n_iter, n_jobs, optimizer.

v1.1.1

• small api-change
• extend progress bar information
• re-enable multiprocessing for new api

v1.0.0

• new API that creates model by function and search space by dict
• enables more flexible usage (e.g. free use of framework, ensembles, nn-structure)
• 100% test coverage

v0.4.2

• performance fixes for bayesian optimization and parallel tempering
• better default parameter for most optimizers
• better implementation for metrics
• add support for catboost
• integration of meta-learn code into hyperactive
• cleanup to avoid similar code

v0.4.1.2

• k-fold-cross validation works with keras models
• a cv of < 1 trains the model on a fraction of the training data and tests on the rest
• better testing and code-coverage
• fix of score and predict method

v0.4.0

• improvement of optimizer class structure
• lower memory usage
• add testing of optimization process
• a lot of clean up und several bug fixes (mostly parallel tempering)