Implementing Estimators#

This page describes how to implement sktime compatible estimators, and how to ensure and test compatibility. There are additional steps for estimators that are contributed to sktime directly.

Implementing an sktime compatible estimator#

The high-level steps to implement sktime compatible estimators are as follows:

  1. identify the type of the estimator: forecaster, classifier, etc

  2. copy the extension template for that kind of estimator to its intended location

  3. complete the extension template

  4. run the sktime test suite and/or the check_estimator utility (see here)

  5. if the test suite highlights bugs or issues, fix them and go to 4

For more guidance on how to implement your own estimator, see this tutorial at pydata on testing interface conformance.

What is my learning task?#

sktime is structured along modules encompassing specific learning tasks, e.g., forecasting or time series classification. For brevity, we define an estimator’s scientific type or “scitype” by the formal learning task that it solves. For example, the scitype of an estimator that solves the forecasting task is “forecaster”. The scitype of an estimator that solves the time series classification task is “time series classifier”.

Estimators for a given scitype should be located in the respective module. The estimator scitypes also map onto the different extension templates found in the extension_templates directory of sktime.

Usually, the scitype of a given estimator is directly determined by what the estimator does. This is also, often, explicitly signposted in publications related to the estimator. For instance, most textbooks mention ARIMA in the context of forecasting, so in that hypothetical situation it makeas sense to consider the “forecaster” template. Then, inspect the template and check whether the methods of the class map clearly onto routines of the estimator. If not, another template might be more appropriate.

The most common point of confusion here is between transformers and other estimator types, since transformers are often used as parts of algorithms of other type.

If unsure, feel free to post your question on one of sktime’s social channels. Don’t panic - it is not uncommon that academic publications are not clear about the type of an estimator, and correct categorization may be difficult even to experts.

What are sktime extension templates?#

Extension templates are convenient “fill-in” templates for implementers of new estimators. They fit into sktime’s unified interface as follows:

  • for each scitype, there is a public user interface, defined by the respective base class. For instance, BaseForecaster defines the fit and predict interfaces for forecasters. All forecasters will implement fit and predict the same way, by inheritance from BaseForecaster. The public interface follows the “strategy” object orientation pattern.

  • for each scitype, there is a private extender interface, defined by the extension contract in the extension template. For instance, the extension template for forecasters explains what to fill in for a concrete forecaster inheriting from BaseForecaster. In most extension templates, users should implement private methods (“inner” methods), e.g., _fit and _predict for forecasters. Boilerplate code rests within the public part of the interface, in fit and predict. The extender interface follows the “template” object orientation pattern.

Extenders familiar with scikit-learn extension should note the following difference to scikit-learn:

the public interface, e.g., fit and predict, is never overridden in sktime (concrete) estimators. Implementation happens in the private, extender sided interface, e.g., _fit and _predict.

This allows to avoid boilerplate replication, such as check_X etc in scikit-learn. This also allows richer boilerplate, such as automated vectorization functionality or input conversion.

How to use sktime extension templates#

To use the sktime extension templates, copy them to the intended location of the estimator. Inside the extension templates, necessary actions are marked with todo. The typical workflow goes through the extension template by searching for todo, and carrying out the action described next to the todo.

Extension templates typically have the following todo:

  • choosing name and parameters for the estimator

  • filling in the __init__: writing parameters to self, calling super’s __init__

  • filling in docstrings of the module and the estimator. This is recommended as early as parameters have been settled on, it tends to be useful as a specification to follow in implementation.

  • filling in the tags for the estimator. Some tags are “capabilities”, i.e., what the estimator can do, e.g., dealing with nans. Other tags determine the format of inputs seen in the “inner” methods _fit etc, these tags are usually called X_inner_mtype or similar. This is useful in case the inner functionality assumes numpy.ndarray, or pandas.DataFrame, and helps avoid conversion boilerplate. The type strings can be found in datatypes.MTYPE_REGISTER. For a tutorial on data type conventions, see examples/AA_datatypes_and_datasets.

  • Filling in the “inner” methods, e.g., _fit and _predict. The docstrings and comments in the extension template should be followed here. The docstrings also describe the guarantees on the inputs to the “inner” methods, which are typically stronger than the guarantees on inputs to the public methods, and determined by values of tags that have been set. For instance, setting the tag y_inner_mtype to pd.DataFrame for a forecaster guarantees that the y seen by _fit will be a pandas.DataFrame, complying with additional data container specifications in sktime (e.g., index types).

  • filling in testing parameters in get_test_params. The selection of parameters should cover major estimator internal case distinctions to achieve good coverage.

Some common caveats, also described in extension template text:

  • __init__ parameters should be written to self and never be changed

  • special case of this: estimator components, i.e., parameters that are estimators, should generally be cloned (via sklearn.clone), and method should be called only on the clones

  • methods should generally avoid side effects on arguments

  • non-state changing methods should not write to self in general

  • typically, implementing get_params and set_params is not needed, since sktime’s BaseEstimator inherits from sklearn’s. Custom get_params, set_params are typically needed only for complex cases only heterogeneous composites, e.g., pipelines with parameters that are nested structures containing estimators.

How to test interface conformance#

For a video tutorial and more examples on the below, please visit our tutorial at pydata.

Using the check_estimator utility#

Usually, the simplest way to test interface conformance with sktime is via the check_estimator methods in the utils.estimator_checks module.

When invoked, this will collect tests in sktime relevant for the estimator type and run them on the estimator.

This can be used for manual debugging in a notebook environment. Example of running the full test suite for NaiveForecaster:

from sktime.utils.estimator_checks import check_estimator
from sktime.forecasting.naive import NaiveForecaster

The check_estimator utility will return, by default, a dict, indexed by test/fixture combination strings, that is, a test name and the fixture combination string in squared brackets. Example: 'test_repr[NaiveForecaster-2]', where test_repr is the test name, and NaiveForecaster-2 the fixture combination string.

Values of the return dict are either the string "PASSED", if the test succeeds, or the exception that the test would raise at failure. check_estimator does not raise exceptions by default, the default is returning them as dictionary values. To raise the exceptions instead, e.g., for debugging, use the argument raise_exceptions=True, which will raise the exceptions instead of returning them as dictionary values. In that case, there will be at most one exception raised, namely the first exception encountered in the test execution order.

To run or exclude certain tests, use the tests_to_run or tests_to_exclude arguments. Values provided should be names of tests (str), or a list of names of tests. Note that test names exclude the part in squared brackets.

Example, running the test test_constructor with all fixtures:

check_estimator(NaiveForecaster, tests_to_run="test_constructor")

{'test_constructor[NaiveForecaster]': 'PASSED'}

To run or exclude certain test-fixture-combinations, use the fixtures_to_run or fixtures_to_exclude arguments. Values provided should be names of test-fixture-combination strings (str), or a list of such. Valid strings are precisely the dictionary keys when using check_estimator with default parameters.

Example, running the test-fixture-combination "test_repr[NaiveForecaster-2]":

check_estimator(NaiveForecaster, fixtures_to_run="test_repr[NaiveForecaster-2]")

{'test_repr[NaiveForecaster-2]': 'PASSED'}

A useful workflow for using check_estimator to debug an estimator is as follows:

  1. Run check_estimator(MyEstimator) to find failing tests

  2. Subset to failing tests or fixtures using fixtures_to_run or tests_to_run

  3. If the failure is not obvious, set raise_exceptions=True to raise the exception and inspecet the traceback.

  4. If the failure is still not clear, use advanced debuggers on the line of code with check_estimator.

Running the test suite in a repository clone#

If the target location of the estimator is within sktime, then the sktime test suite can be run instead. The sktime test suite (and CI/CD) is pytest based, pytest will automatically collect all estimators of a certain type and tests applying for a given estimator.

For an overview of the testing framework, see the “testing framework” documentation. Generic interface conformance tests are contained in the classes TestAllEstimators, TestAllForecasters, and so on. pytest test-fixture-strings for an estimator EstimatorName will always contain EstimatorName as a substring, and are identical with the test-fixture-strings returned by check_estimator.

To run tests only for a given estimator from the console, the command pytest -k "EstimatorName" can be used. This will typically have the same effect as using check_estimator(EstimatorName), only via direct pytest call. When using Visual Studio Code or pycharm, tests can also be sub-set using GUI filter functionality - for this, refer to the respecetive IDE documentation on test integration.

To identify codebase locations of tests applying to a specific estimator, a quick approach is searching the codebase for test strings produced by check_estimator, preceded by def (for function/method definition).

Testing within a third party extension package#

For third party extension packages to sktime (open or closed), or third party modules that aim for interface compliance with sktime, the sktime test suite can be imported and extended in two ways:

  • importing check_estimator, this will carry out the tests defined in sktime

  • importing test classes, e.g., test_all_estimators.TestAllEstimators or test_all_forecasters.TestAllForecasters. The imports will be discovered directly by pytest. The test suite also be extended by inheriting from the test classes.

Adding an sktime compatible estimator to sktime#

When adding an sktime compatible estimator to sktime itself, a number of additional things need to be done:

  • ensure that code also meets sktime's documentation standards.

  • add the estimator to the sktime API reference. This is done by adding a reference to the estimator in the correct rst file inside docs/source/api_reference.

  • authors of the estimator should add themselves to the "authors" and "maintainers" tag of the estimator, as owners of the contributed estimator.

  • if the estimator relies on soft dependencies, or adds new soft dependencies, the steps in the “dependencies” developer guide should be followed

  • ensure that the estimator passes the entire local test suite of sktime, with the estimator in its target location. To run tests only for the estimator, the command pytest -k "EstimatorName" can be used (or vs code GUI filter functionality)

  • ensure that test parameters in get_test_params are chosen such that runtime of estimator specific tests remains in the seconds order on sktime remote CI/CD

Don’t panic - when contributing to sktime, core developers will give helpful pointers on the above in their PR reviews.

It is recommended to open a draft PR to get feedback early.

Estimators dependent on cython#

To add an estimator to sktime that depends on cython, the following additional steps are needed:

  • all cython code should be present in a separate package on pypi and/or conda-forge. No cython dependent code should be added directly to sktime. Below, we call this separate package home-package, for simplicity of reference.

  • In home-package, it is recommended to test the estimator via check_estimator, on the same test matrix as sktime: all supported python versions; MacOS, Linux, Windows.

  • In sktime, an interface to the algorithm should be added. This can be a simple import from home-package, if the algorithm in home-package already passes check_estimator.

  • Alternatively, the algorithm can be interfaced via a delegator as a delegate, tags and method overrides can be added in the delegator. See, e.g., MrSQM for this.

  • For the sktime interface, the requires_cython tag should be set to True, and the python_dependencies tag should be set to the string "home-package".

If all has been setup correctly, the estimator will be tested in sktime by the CI element test-cython-estimators. Note that this CI element does not cover the full test matrix of python version and operating systems, this should be done in the upstream package.