What's New
==========


v0.4.0 (June 18, 2020)
-------------------------

Breaking changes
~~~~~~~~~~~~~~~~

- Python <3.6 is no longer officially supported. The package might still work,
  but we don't test against these versions anymore.


Enhancements
~~~~~~~~~~~~

- The package can now be installed via conda::

    conda install -c phausamann -c conda-forge sklearn-xarray



v0.3.0 (November 5, 2018)
-------------------------

Breaking changes
~~~~~~~~~~~~~~~~

- ``wrap`` now returns a new class ``CompatEstimatorWrapper`` when
  ``compat=True``.
- The standard ``EstimatorWrapper`` directly reflects the parameters of the
  underlying estimator as instance attributes, regardless of the value of
  ``compat`` (which is deprecated and has no effect).

Enhancements
~~~~~~~~~~~~

- ``EstimatorWrapper`` now directly reflects both the parameters and the
  fitted attributes (e.g. ``mean_``) of the underlying estimator. The
  ``estimator`` attribute is still an instance of the actual estimator but is
  treated mostly as just the ``type`` of the instance (It's not stored as
  the type for compatibility with ``clone``).
- Added the ``CompatEstimatorWrapper`` which acts like a standard sklearn
  estimator (with the wrapped estimator as nested) and does not
  present the attributes of the underlying estimator.
- Added ``inverse_transform`` to ``preprocessing.Concatenator``.

Bug fixes
~~~~~~~~~

- Fixed failing tests with sklearn 0.20.


v0.2.0 (April 9, 2018)
----------------------

Breaking changes
~~~~~~~~~~~~~~~~
- ``wrap`` now returns a decorated ``EstimatorWrapper`` instead of an
  estimator-specific wrapper class.
- Removed the ``common.decorators`` module, because the decorated
  estimators could not be pickled and therefore didn't pass the usual sklearn
  estimator checks.
- Removed the ``dataset`` and ``dataarray`` modules. Wrappers have
  to be directly imported from ``sklearn_xarray``.
- Removed the ``data`` module (now called ``datasets``).


Enhancements
~~~~~~~~~~~~

- Added wrappers for ``fit_transform``, ``partial_fit``, ``predict_proba``,
  ``predict_log_proba`` and ``decision_function``.


v0.1.4 (March 15, 2018)
-----------------------

Enhancements
~~~~~~~~~~~~

- ``preprocessing.Transposer`` now also accepts a subset of ``X.dims`` for the
  ``order`` parameter.
- ``preprocessing.Splitter`` and ``preprocessing.Segmenter`` now accept an
  ``axis`` argument that specifies where to insert the new dimension.
- Huge performance improvements for ``preprocessing.Segmenter`` by using
  ``numpy.lib.stride_tricks.as_strided`` instead of a loop. The
  general-purpose backend for segmenting can be found in
  ``utils.segment_array``.

Deprecations
~~~~~~~~~~~~

- The ``data`` module containing different example datasets is being renamed
  to ``datasets`` according to the scikit-learn standards. Since the
  ``dataset`` module will be removed, there will no longer be confusion due
  to similar naming.


v0.1.3 (January 9, 2018)
------------------------

Enhancements
~~~~~~~~~~~~

The wrapper now passes the DataArray's ``data`` attribute to the wrapped
estimator, making it possible to wrap estimators from dask-ml_ and use
dask-backed DataArrays and Datasets as inputs.

.. _dask-ml: http://dask-ml.readthedocs.io/en/latest/index.html


v0.1.2 (December 10, 2017)
--------------------------

Enhancements
~~~~~~~~~~~~

The wrapping mechanism has been changed to work with both DataArrays and
Datasets. From now on, you can use ``from sklearn_xarray import wrap`` which
will automatically determine the type of xarray object when calling ``fit``.
Note that a wrapper fitted for DataArrays cannot be used for Datasets and
vice-versa.

The wrappers now also support passing an estimator type rather than an
instance and passing the parameters of the wrapped estimator directly to the
wrapper.