Skip to content

CHANGELOG.md

Latest commit

 

History

History
177 lines (146 loc) · 7.11 KB

CHANGELOG.md

File metadata and controls

177 lines (146 loc) · 7.11 KB

v.0.12.1

  • Dataset features can now be easily accessed with the property dataset.features
  • model.test_estimators with CV will keep using CV if it's refitting the best estimator.
  • Result now has a .parameters attribute to show what parameters generated the result
  • Switch to pyproject.toml for project metadata
  • Updated documentation to use california housing dataset instead of boston
  • Updated documentation to remove deprecated parameters from estimators

v0.12.0

  • Permutation importance and Feature importance are now two different plotting methods.
  • Model.test_estimators now takes a feature_pipeline argument
  • Fixed a bug where FillNA did not create a _is_na column if the column didn't have a missing value
  • Implemented Bayesian Search for hyperparameter optimization
  • Added a read_file convenience method to FileDataset to read
  • Fixed a bug where copy_to failed between two instances of Sqlite based SQLDatasets
  • Fixed a bug where ClassificationVisualize.confusion_matrix would fail on multi-class problems due to wrong defaults
  • Added repr to demodataset
  • Lift curve now can plot multi-class
  • Precision-Recall curve can now plot multi-class
  • ROC AUC curve can now plot multi-class
  • Fixed Binner to have a default value
  • Fixed FuncTransform to have a default value
  • load_estimator now uses default storage if nothing is passed
  • Model.bayessearc is now Model.bayesiansearch
  • Added target_feature_distributionto Dataset.plot

v0.11.0

  • Added load_demo_dataset function
  • If the dataset has no train set score_estimator will now run create_train_test with default configurations
  • Model.make_prediction now takes a threshold argument when making a binary classification
  • All ML-tooling logging messages now go to stdout instead of stderr
  • Can pass a feature pipeline to Model which will then automatically generate a combined feature_pipeline + estimator Pipeline
  • Can pass a feature pipeline to Dataset.plot methods, to apply preprocessing before visualization
  • New config implementation. If you need to reset the configuration, you should use Model.config.reset_config()

v0.10.3

  • Fixed typehints in Dataset
  • Dataset.create_train_test now takes a boolean stratify parameter.
  • Added default local filestorage when using save_estimator
  • The dataframe returned by .make_predictionnow labels the columns in a more human friendly manner
  • Dataset now verifies that load_training_data and load_prediction_data do not return empty
  • Added a missing data visualization to Dataset.plot
  • FillNA now accepts a is_nanflag which adds a flag indicating that a value was missing
  • Model.make_prediction now accepts a use_cacheflag to score everything in cached .x
  • Added a new Transformer: RareFeatureEncoder

v0.10.2

  • Fixed type inferences from data to sql in _load_data
  • Added idx arg to load_prediction_data abstract method in SQLDataset
  • Added caching of loaded data in SQLDataset

v0.10.1

  • Added .copy_to functionality to SQLDataset and FileDataset, allowing copying between datasets

v0.10.0

  • Bug fix for calculating feature importance when passing large amounts of data
  • Bug fix when using default metric in test_estimators
  • Bug fix when gridsearching, only applying last change
  • Add nicer error message when passing incorrect dtypes to FillNA
  • Storage .save method now only takes filename as parameter
  • Handles storage loading of paths outputted from the Storage .get_list method
  • Handles case when Dataset does not have a y value
  • Added plot_learning_curve and corresponding result.plot.learning_curve
  • Added plot_validation_curve and corresponding result.plot.validation_curve
  • Replaced permutation_importance with scikit-learn's implementation
  • Added target_correlation plots to Dataset.plot

v0.9.2

  • Bug fix for logging when feature unions (DFFeatureUnion) had tuples

v0.9.1

  • Hot fix python version to 3.7

v0.9.0

  • Breaking change - Model methods load_estimator and save_estimator now takes a Storage class that defines how and where to store estimators.
  • Added the ability to declare that a saved model should be a production estimator.
  • Added corresponding .load_production_estimator to Model

v0.8.1

  • Removed gitpython as a dependency

v0.8.0

  • Replaced custom feature permutation importance with sklearns implementation from v0.22
  • Breaking change - Dataset is now a separate object that has to be instantiated outside Modeldata
  • Breaking change - ModelData is now renamed to Model
  • Added new properties is_estimator and is_regressor which checks what type of estimator is used

v0.7.1

  • Joblib is now a dependency, instead of being vendored with scikit-learn
  • Updated requirements

v0.7.0

  • Breaking change - BaseClassModel renamed to ModelData.
  • Breaking change - model renamed to estimator
  • Added Precision-Recall Curve
  • Added option to give custom file name to .save_estimator()
  • Instantiating with estimator is now optional - set estimator later using .init_estimator
  • We have a logo! Added ML Tooling logo to docs

v0.6.2

Bugfixes

  • Now issues a warning when git is not installed.

v0.6.1

Bugfixes

  • Data for a class is changed from instance variable to class variable
  • Grid search only copies data to workers once and reuses them across grid and folds.
  • The Data Class now takes a random seed which it will receive from the BaseClass
  • Disabled mem-maping in feature importance
  • Added license file to package
  • Updated requirements

v0.6.0

Features

  • Feature importances changed to use permutation instead of built-in for better estimates.

Bugfixes

  • .train_estimator will now reset the result attribute to None, in order to prevent users from mistakenly assuming the result is from the training
  • Fixed bug in lift_score when using dataframes
  • Fixed bug when training model and then scoring model
  • Fixed bug where users could not save models if no result had been created, as would happen if the user only called .train_estimator before saving.
  • Default_metric is now the same metric as the one specified for the model in .config
  • Each class inheriting from ModelData has an individual config
  • Changed get_scorer_func to wrap sklearn's get_scorer
  • Fixed bug when gridsearching twice

v0.5.0

  • Added Binarize Transformer
  • Added ability to use keywords in FuncTransformer
  • .predict now returns a dataframe indexed on input
  • Updated dependencies

v0.4.0

  • Added gridsearch method to BaseClass. Gridsearch your model and return a list of results for inspection
  • Added ResultGroup - any method that returns a list of results now returns a ResultGroup instead.
  • Added logging
  • Added ability to record runs as yaml

v0.3.2

  • Another bugfix release

v0.3.1

  • Fixed bug that prevented DFRowFunc from pickling properly

v0.3.0

  • Added DFRowFunc Transformer
  • Updated FillNA to handle categorical values
  • Allow user to choose whether score_model uses cv or not

v0.2.3

  • Plot_feature_importance now takes a top_n and bottom_n argument

v0.2.2

  • Fix for error in setup wheels

v0.1.3

  • Implemented new FillNA Transformer

v0.1.2

  • Refactored to use flat structure

v0.1.1

  • Renamed project to ml_tooling

v0.1.0

  • Initial release