Release Notes

Version 23.0.0.post1

Fix scikit-learn to version <1.3, because GAMA uses scikit-learn internals that were changed from 1.4.

Version 23.0.0

Highlights:

  • We (temporarily) removed the Dashboard, it had not been maintained and is best be rebuilt from the ground up.

  • The config hyperparameter is now called search_space.

  • Adds back the preset hyperparameter to all GAMA estimators which let you specify the goal of your AutoML execution. Currently simple can be specified to create a simple model, and performance can be used to generate the best possible model. It is still possible to manually set the search and post processing methods.

Maintenance:

  • Bump minimum Python version to Python 3.9, in line with NEP 29.

  • Bump several library dependencies.

  • Switch from flake8 to ruff.

Version 22.0.0

Maintenance:
  • Adopt NEP 29:

    • Drop support for Python 3.6 and 3.7

    • Add support for Python 3.10

  • Use pyproject.toml instead of setup.py

  • Transition to Github CI, including:

    • pytest on pull request

    • pytest CRON job with prerelease versions

    • doc build and deploy

    • pre-commit check

    • check for changelog

    • easy release to pypi from github workflow

  • Small changes to avoid FutureWarnings and/or DeprecationWarnings.

Bugfixes:

  • #137: raise an output if output_directory is non-empty.

  • #174: Fix an issue where GAMA might freeze in when ending search.

Features:

  • ASHA resources can now also be specified as fraction of the dataset size by using a float in (0, 1].

Changes:
  • #138: Instead of subsampling the whole dataset before doing CV, the same test set is used across rungs and only

    the training folds are subsampled. This makes performance comparable across rungs.

  • AsyncEA will try to delay creating new individuals as long as possible.

  • AsyncEA will no longer create offspring from pipelines with failed evaluations (e.g., timeout, invalid configuration).

Version 21.0.1

News:

  • We have a logo! Check out our README, or imageslogos for several renders.

Bugfixes:

  • Check if max_pipeline_length is compatible with the search space, i.e. if there are no preprocessing steps in the search space it is set to 1 and raises an error if already set to a value greater than 1.

  • mut_replace_primitive mutation is no longer considered if there are no suitable replacements for any primitive in the individual.

  • Setting store to “all”.

  • Setting n_jobs to -1 once more correctly creates one subprocess per core.

  • No longer leave zombie processes when fit is called more than once per process.

Contributors (alphabetical order):

  • @Bilgecelik

  • @PGijsbers

Version 21.0.0

Features:

  • The warm_start parameter of fit() had a slight behavior change (see docs).

  • Fit can now be called more than once. It will use the same time constraint as the first fit call.

Bugfixes:

  • TargetEncoder is no longer used for Classification tasks, since there is a bug that prevents it from working as intended.

Version 20.2.2

Maintenance:

  • Minor refactoring

  • Updated and expanded documentation

Bugfixes:

  • Avoid a cache-related error when using ASHA search.

Version 20.2.1

Changes:
# 24: Changes to logging

The structure of the log file(s) have changed. The goal is to make the log files easier to use, by making them easier to read and easier to extend write behavior. There will now be three log files, one which contains just evaluation data, one which contains progress data, and one which contains resource usage data. For more information see Logging in the technical guide.

Features:
# 66: csv files are now supported.

Call fit_arff is now fit_from_file which accepts both arff and csv files. The CLI interface and Gama Dashboard also allow for csv files.

# 92: You can specify a memory limit through max_memory_mb hyperparameter.

GAMA does not guarantee it will not violate the constraint, but violations should be infrequent and minor. Feel free to open an issue if you experience a violation which does not minor.

Version 20.2.0

Features:

# 70: Exported code is now auto-formatted with Black. - The Dashboard now has an Analysis tab that allows you to load old log files. - The Dashboard Home tab allows you to view the data set. - The Dashboard Home tab allows you to select a target column. - Pipelines and estimators are now cached on disk again. - KNN and PolynomialFeatures are now dynamically disabled based on dataset size.

Maintenance:

  • Docs are automatically built and deployed on a commit to master and develop.

  • Pre-commit configuration added to check formatting, style and type hints on commit.

  • Black codestyle adapted, most drastic change is line length from 120 to 88.

  • Coverage increased by removing unused code, updating configuration, adding tests.

  • Memory usage of all GAMA’s processes is logged.

Version 20.1.0

Features:

  • Encoding of ARFF files may now be specified with the encoding parameter in {fit/predict/score}_arff calls.

  • Set max_pipeline_length on initialization to impose a maximum number of steps in your pipelines.

Bugfixes:

  • Reading ARFF markers (such as @data and @attribute) is now correctly case insensitive.

Maintenance:

  • Evaluation results are no longer saved to disk but kept in memory only. Consequently the cache_dir hyperparameter has been removed.

Changes:

  • Pipelines fitted during search are now used in the ensemble, instead of retraining the pipeline.

  • Ordinal Encoding and One Hot Encoding are now applied outside of 5-fold CV. This is for computational reasons, as all levels of a categorical variable are known this shouldn’t make a difference.

Version 20.0.0

Features:
#65 GAMA Command Line Interface:

Allows users to start GAMA from the command line. Requires data to be formatted in ARFF.

#69 Code export:

Export Python code that sets up the machine learning pipeline found with AutoML.

#71 GAMA Dashboard:

First steps to providing a user interface for GAMA. It allows users to start GAMA AutoML through a webapp built with Dash, and monitor the performance of the search as it is executed.

Bugfixes:

#68: Only add categorical encoding steps if (non-binary) categorical data is present.

Maintenance:

#67: Selection now takes crowding distance into account (again). #68: n_jobs will now default to use half of available cores. #68: Updates given about the Pareto front now include the pipeline structure. #70: Versioning now YY.Minor.Micro

Version 19.11.2

Bugfixes:

  • predict_proba_arff now also accepts a target_column as expected from the previous update.

Version 19.11.1

Features:

  • gama.__version__ can now be used to retrieve gama’s version.

  • fit_arff, score_arff and predict_arff now accept a target_column parameter to specify the target. If left unset, the last column of the ARFF file is assumed to be the target column.

Bugfixes:

  • fit(x, y) may now be called with y as (N,1) array.

  • ensemble post-processing is now compatible with non-zero indexed class labels

Maintenance:

  • __version__.py is now the only place with hard-coded version.

Version 19.11.0

Accidentally released without updates.

Version 19.08.0

  • Prototype dash app for visualizing GAMA logs.

  • Easy switching between search algorithms

Version 0.1.0

First GAMA release.