Support newest matplotlib release #700
Conversation
Empty distributions are plottable in mpl < 1.4.0. In 1.4.0, a ValueError is raised. This has been fixed in mpl 1.4.0-dev (see matplotlib/matplotlib#3571). In order for skbio.draw.boxplots to support empty distributions across mpl versions, empty distributions are replaced with [np.nan]. See pandas-dev/pandas#8382 and pandas-dev/pandas#8240 for details.
Conflicts: setup.py skbio/draw/_distributions.py
| # pre-1.4.0 version and whatever the latest version is. | ||
| - PYTHON_VERSION=3.4 NUMPY_VERSION= MATPLOTLIB_VERSION= | ||
| - PYTHON_VERSION=2.7 NUMPY_VERSION= MATPLOTLIB_VERSION= WITH_DOCTEST=True USE_CYTHON=TRUE | ||
| - PYTHON_VERSION=2.7 NUMPY_VERSION='=1.7' MATPLOTLIB_VERSION='=1.3.1' WITH_DOCTEST=True |
There was a problem hiding this comment.
It'd be more consistent and cleaner to set FOO_VERSION=x.y and keep the = in the conda create command.
There was a problem hiding this comment.
Ah, didn't think that'd work for the case where we don't want to specify a version (i.e., to test against the latest version). Will update, thanks!
There was a problem hiding this comment.
This doesn't work (see https://travis-ci.org/biocore/scikit-bio/jobs/37241602). Reverting back to what I had before.
There was a problem hiding this comment.
Ah, sorry! My bad, didn't realize how you were getting the latest version instead of hardcoding it. Cool!
There was a problem hiding this comment.
No worries, I wish there was a better way to accomplish this 😞
|
This looks good, thanks @jairideout! |
Support newest matplotlib release
The newest matplotlib release (1.4.0) had its boxplot functionality largely rewritten/refactored. There were a few unintended API/behavior changes that affect the boxplotting functionality in skbio (
skbio.draw.boxplots). The most noticeable change is that aValueErroris raised if an empty distribution is provided; in the past, this type of input was supported.This bug has been fixed in the latest development version of matplotlib (matplotlib/matplotlib#3571). It was first noticed in pandas (pandas-dev/pandas#8382) and they worked around this issue by passing
np.array([np.nan])instead of an empty sequence (pandas-dev/pandas#8240). We use their approach here in order support plotting empty distributions in pre- and post-1.4.0 matplotlib releases. If we decide to only support matplotlib > 1.4.0 at some point in the future, this NaN hack can be removed.Some of the existing boxplot test code had to be updated to work in either matplotlib version; this was due to changes in the fliers returned by
pyplot.boxplot(matplotlib/matplotlib#3544). I also added (very ugly) test code to test that the NaN hack appears to be working; this will hopefully catch changes to matplotlib's boxplot functionality in the future if the hack stops working as intended.I updated Travis to test against matplotlib < 1.4.0 and whatever the latest release is, in order to make sure that our code works with either version.
Fixes #649.