Model Tuner 0.0.19a

• Requirements updated again to make compatible with google colab out of the box.
• Bug in fit() method where best_params wasn't defined if we didn't specify a score
• Threshold bug now actually fixed. Specificity and other metrics should reflect this. (Defaults to 0.5 if optimal_threshold is not specified).

Model Tuner 0.0.18a

• Updated requirements to include numpy versions <1.26 for Python 3.8-3.11.

This should stop a rerun occurring when using the library on a google colab.

Model Tuner 0.0.17a

Major fixes:

• Verbosity variable is now popped from the parameters before the fit
• Bug with Column Transformer early stopping fixed (valid set is now transformed correctly)
• Return metrics now has a consistent naming convention
• report_model_metrics is now using the correct threshold in all cases
• Default values updated for train_val_test_split
• tune_threshold_Fbeta is now called with the correct number of parameters in all cases
• Requirements updates: XGBoost updated to 2.1.2 for later Python versions.

Minor changes:

• help(model_tuner) should now be correctly formatted in google colab

Model Tuner 0.0.16a

Version 0.0.16a

• Custom pipeline steps now updated (our pipeline usage has been completely changed and should now order itself and support non named steps) always ensures correct order
• This fixed multiple other issues that were occuring to do with logging of imbalanced learn
• Reporting model metrics now works.
• AutoKeras code deprecated and removed.
• KFold bug introduced because of CatBoost. This has now been fixed.
• Pretty print of pipeline.
• Boosting variable has been renamed.
• Version constraints have been updated and refactored.
• tune_threshold_Fbeta has been cleaned up to remove unused parameters.
• train_val_test unnecessary self removed and taken outside of class method.
• deprecated setup.py in favor of pyproject.toml per forthcoming pip25 update.

Model Tuner 0.0.15a

Version 0.0.15a

Contains all previous fixes relating to:

• CatBoost support (early stopping, and support involving resetting estimators).
• Pipeline steps now support hyperparameter tuning of the resamplers (SMOTE, ADASYN, etc.).
• Removed older implementations of impute and scaling and moved onto supporting only custom pipeline_steps.
• Fixed bugs in stratification with regards to length mismatch of dependent variable when using column names to stratify.
• Cleaned a removed multiple lines of unused code and unused initialisation parameters.

Model Tuner 0.0.14a

Version 0.0.14a

In previous versions, the train_val_test_split method allowed for stratification either by y (stratify_y) or by specified columns (stratify_cols), but not both at the same time. There are use cases where stratification by both the target variable (y) and specific columns is necessary to ensure a balanced and representative split across different data segments.

Enhancement

Modified the train_val_test_split method to support simultaneous stratification by both stratify_y and stratify_cols. This was inside the method achieved by implementing the following logic that ensures both y and the specified columns are considered during the stratification process.

stratify_key = pd.concat([X[stratify_cols], y], axis=1)

strat_key_val_test = pd.concat(
    [X_valid_test[stratify_cols], y_valid_test], axis=1
)

Model Tuner 0.0.13a

Version 0.0.13a

• Updated bootstrapper
• evaluate_bootstrap_metrics
• Added notebooks/xgb_early_bootstrap_test.py to test it
• Updated requirements.txt file for dev testing
• Fixed sampling error on low number of samples inside bootstrapper

Model Tuner 0.0.12a

• Xgboost bug fixes
• Zenodo updates
• Pickle model fixes with np import
• ADASYN and SMOTE fix with no fit happening when calibrating

model_tuner 0.0.11a

• updated readme for PyPI
• previous version not saved on setup; re-release to 0.0.11a

model_tuner 0.0.10a

• updated readme for PyPI