Overview of basic unit test framework for sktime

References issues

#246, #25

All estimators

Tests common to all estimator objects in sktime, adapted list from scikit-learn's estimator_checks:

• check has base API attributes
• check inherits from BaseEstimator and one of the task-specific estimators
• check ("predict", "transform", "decision_function", "predict_proba", "inverse_transform") does not change fitted parameters
• check fit does not overwrite hyper-parameters (also add checks for other methods like predict, transform etc)
• check not fitted error for ("predict", "transform", "decision_function", "predict_proba", "inverse_transform")
• check fit updates state ("is_fitted", "_is_fitted")
• check fit returns self
• check fit idempotent
• check set_params returns self
• check repr
• check clone
• test no logic/interaction in constructor
• check get-params invariance (use sklearn function: check_get_params_invariance)
• check set params (use sklearn function: check_set_params)
• check persistence via pickle
• check pipeline invariance
• different data types in input data container (empty, nan, inf)
• check corner cases for fit/predict/transform (1 sample, 1 timepoint, 1 column)

Series-as-features predictors (regressors/classifiers/clustering algorithms)

• check inputs
• check multivariate estimators, create tags _tags = ("multivariate") as class variable
• check fit full vs fit by batch invariance
• check fit and score always accept y as second input argument
• check input target (missing, single lable, zero variance)
• check predict output
• check score method

Series-as-features transformers

• input data and transformed data has the same index
• check multivariate estimators

Forecasters

• check inputs
• check input target warnings (zero variance)
• check predict output
• test score
• test predict in-sample
• test predict out-of-sample
• test update predict
• test fh
• test oh

Ignored scikit-learn tests

• sample weight tests
• class weight tests