Suggesting improvements for bar plot orientation behaviour. #7994

Bug report

Bug summary

Guided by this tutorial example user could expect that the same applies for any graphing function with the keyword orientation. This isn't the default case when dealing with bar. Specifying orientation keyword does not produce a rotated bar graph without manually setting the required parameters to avoid an error. To achieve the desired result barh has to be used.

Code for reproduction

import matplotlib.pyplot as plt

#import numpy as np

#N = 5
#img = np.random.randn(N, N)
#xs = img.sum(axis=0)
#ys = img.sum(axis=1)
#bins = np.arange(0, N, 1)

img = [[-0.13448386, -0.92189708, -0.3883107 ,  1.83910067,  0.46092219],
       [-0.29775299,  0.4623531 , -0.53428658,  0.82103232, -0.66423673],
       [ 1.4559248 , -0.74152634, -0.10229999,  1.12182625, -0.20793333],
       [-1.38758209,  0.61128182, -1.34506956, -0.5511364 ,  0.50791828],
       [ 1.29593156,  0.28209814,  1.60556993,  0.71651822,  0.11639   ]]

xs = [ 0.93203742, -0.30769036, -0.7643969 ,  3.94734105,  0.21306041]
ys = [ 0.85533123, -0.21289089,  1.5259914 , -2.16458795,  4.01650784]
bins = [0, 1, 2, 3, 4]

fig = plt.figure()
plt.subplot(221)
plt.bar(left=None, height=0.8, bottom=bins, width=ys,
          orientation="horizontal")

plt.subplot(222)
plt.bar(bins, ys, orientation="horizontal")

plt.subplot(223)
plt.barh(bins, ys)

plt.subplot(224)
plt.bar(bins, xs)

plt.show()

Actual outcome

Current output for the above code will fail with an error.

Traceback (most recent call last):
  File "/usr/local/lib/python2.7/dist-packages/matplotlib/axes/test.py", line 18, in <module>
    plt.bar(bins, ys, orientation="horizontal")
  File "/usr/local/lib/python2.7/dist-packages/matplotlib/pyplot.py", line 2705, in bar
    **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/matplotlib/__init__.py", line 1892, in inner
    return func(ax, *args, **kwargs)
  File "/usr/local/lib/python2.7/dist-packages/matplotlib/axes/_axes.py", line 2075, in bar
    "be length %d or scalar" % nbars)
ValueError: incompatible sizes: argument 'left' must be length 1 or scalar

Expected outcome

• First graph (subplot 221, top left) will not produce an error and represents the currently correct usage with orientation horizontal. It's here to show that this isn't a bug, just weird usage (API).
• Second graph (subplot 222, top right) will produce said error. The error happens because the default parameter order is bar(self, left, height, width=0.8, bottom=None, **kwargs): so left and height are set by default instead of being dependent on orientation keyword in which case sometimes bottom and width parameters should be set instead, and height should become width. Error is raised when the method bar checks data for consistency but doesn't find any.
• Third graph (subplot 223, bottom left) output should be the same as first two - nothing should be broken for backward compatibility issues as well as the fact that histogram with bar option and horizontal orientation secretly calls barh for drawing.
• Fourth graph (subplot 224, bottom right) is the vertical orientation and should display correctly after the changes.
  

Proposed change

In the case when bar is called with left and horizontal orientation:

• values in bottom and left parameters should be exchanged
• values in width and height should be exchanged

Changes affect the axes/_axes.py file line no. 2034 where I propose that the values are manually exchanged, as in the code below, between the parameters before _process_unit_info is called, that is immediately after elif orientation == "horizontal" line:

if _left is not None:
    _bottom, bottom, _left, left = _left, left, _bottom, bottom
    width, height = height, width

Once this is implemented the listed examples above produce the expected result.

As a final thought I believe that solution is a bit "hackish" because it makes it harder to track data in the code for future devs/users as the switch happens relatively late inside bar. On the other hand its more user friendly because it reduces the amount of knowledge users have to have about how mpl works on the inside and makes it possible to more intuitively work with the mpl's API. In practice I believe code should be refactored to remove barh and standardize the orientation keyword to work even in the cases of line plots or scatter plots. In principle this would consist of just a function that would switch the x and y data parameters. This would not be a very friendly change, and a bit of an overkill, as old code depending on barh would not work anyomre and at the same time a lot of mpl's code already seems to depend on the bar barh difference to work properly.

Additionally, I have no idea if this change breaks something else, but so far it dosen't seem to have hurt my installation very much.

Matplotlib version

• Matplotlib version 2.0.0, installed from pip, numpy version 1.7.1
• Python 2.7.6

Sorry if this was an issue before, I looked but found none.