All Patsy releases are archived at Zenodo:
- The Python 3.6.7 and 3.7.1 point releases changed the standard tokenizer module in a way that broke patsy. Updated patsy to work with these point releases. (See #131 for details.)
- On Python 2, accept
unicodestrings containing only ASCII characters as valid formula descriptions in the high-level formula API (
dmatrix()and friends). This is intended as a convenience for people using Python 2 with
from __future__ import unicode_literals. (See Python 2 versus Python 3.)
longas a valid integer type in the new
DesignInfoclasses. In particular this fixes errors that arise on 64-bit Windows builds (where
ValueError: For numerical factors, num_columns must be an int.
- Fix deprecation warnings encountered with numpy 1.10
ModelDesc.from_formula()no longer take an
design_matrix_builders()function and the
factor_protocol.memorize_passes_needed()method now require an
eval_envas an additional argument.
DesignInfoconstructor’s arguments have totally changed. In addition to the changes needed to support the new features below, we no longer support “shim” DesignInfo objects that have non-trivial term specifications. This was only included in the first place to provide a compatibility hook for competing formula libraries; four years later, no such libraries have shown up. If one does, we can re-add it, but I’m not going to bother maintaining it in the mean time…
- Dropped support for Python 3.2.
- Patsy now supports Pandas’s new (version 0.15 or later) categorical objects.
- Formulas (or more precisely,
EvalFactorobjects) now only keep a reference to the variables required from their environment instead of the whole environment where the formula was defined. (Thanks to Christian Hudon.)
DesignInfohas new attributes
DesignInfo.term_codingswhich provide detailed metadata about how each factor and term is encoded.
- As a result of the above changes, the split between
DesignMatrixBuilderis no longer necessary;
DesignMatrixBuilerhas been eliminated. So for example,
design_matrix_builders()now returns a list of
DesignInfoobjects, and you can now pass
DesignInfoobjects directly to any function for building design matrices. For compatibility,
DesignInfocontinues to provide
.design_infoattributes, so that old code should continue to work; however, these attributes are deprecated.
- Ensured that attempting to pickle most Patsy objects raises an error. This has never been supported, and the interesting cases failed in any case, but now we’re taking a more systematic approach. (Soon we will add real, supported pickling support.)
- Fixed a bug when running under
New stateful transforms for computing natural and cylic cubic splines with constraints, and tensor spline bases with constraints. (Thanks to @broessli and GDF Suez for contributing this code.)
Dropped support for Python 2.5 and earlier.
Switched to using a single source tree for both Python 2 and Python 3.
Added a fast-path to skip NA detection for inputs with boolean dtypes (thanks to Matt Davis for patch).
Incompatible change: Sometimes when building a design matrix for a formula that does not depend on the data in any way, like
"1 ~ 1", we have no way to determine how many rows the resulting matrix should have. In previous versions of patsy, when this occurred we simply returned a matrix with 1 row. In 0.3.0+, we instead refuse to guess, and raise an error.
Note that because of the next change listed, this situation occurs less frequently in 0.3.0 than in previous versions.
build_design_matrices()(or derived functions like
dmatrices()) is a
pandas.DataFrame, then we now check its number of rows and index, and insist that the output design matrices match. This also means that if
datais a DataFrame, then the error described in the first bullet above cannot occur – we will simply return a column of 1s that is the same size as the input dataframe.
Worked around some more limitations in py2exe/py2app and friends.
- Fixed a nasty bug in missing value handling where, if missing values
dmatrix(..., result_type="dataframe")would always crash, and
dmatrices("y ~ 1")would produce left- and right-hand side matrices that had different numbers of rows. (As far as I can tell, this bug could not possibly cause incorrect results, only crashes, since it always involved the creation of matrices with incommensurate shapes. Therefore there is no need to worry about the accuracy of any analyses that were successfully performed with v0.2.0.)
patsy/__init__.pyto work around limitations in py2exe/py2app/etc.
- The lowest officially supported Python version is now 2.5. So far as I know everything still works with Python 2.4, but as everyone else has continued to drop support for 2.4, testing on 2.4 has become so much trouble that I’ve given up.
- New support for automatically detecting and (optionally) removing
missing values (see
- New stateful transform for B-spline regression:
bs(). (Requires scipy.)
- Added a core API to make it possible to run predictions on only a
subset of model terms. (This is particularly useful for
e.g. plotting the isolated effect of a single fitted spline term.)
LookupFactornow allows users to mark variables as categorical directly.
pandas.Categoricalobjects are now recognized as representing categorical data and handled appropriately.
- Better error reporting for exceptions raised by user code inside formulas. We now, whenever possible, tag the generated exception with information about which factor’s code raised it, and use this information to give better error reporting.
EvalEnvironment.capture()now takes a reference argument, to make it easier to implement new
Other: miscellaneous doc improvements and bug fixes.
First public release.