Fix FixedFormatter

[mdboom]

In investigating #4301, I discovered that when you set explicit tick labels, the value display in the toolbar doesn't work. Strangely, it seems to have been this way at least since 1.1 (before that, I can't compile due to libpng API changes).

To reproduce:

import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt

f, ax = plt.subplots(figsize=(9, 9))
plt.plot(np.arange(100))
plt.gca().set_xticklabels(['foo %s' % i for i in range(100)])
plt.show()

Looking at the FixedFormatter code, it seems that it's supposed to be keying the values from a kwarg pos, but I can't see anywhere in the code where pos is being passed...

Does this change make sense? I think probably not, but it's broken as it stands in any event...

Cc: @efiring

[WeatherGod]

hmm, dug a bit deeper into the format_coord() method. There is a failback mechanism if calling fmt_xdata() raises a TypeError that uses the axis' major formatter. That might explain this a bit.

[mdboom]

@WeatherGod: How so?

[WeatherGod]

Because the default formatter is ScalarFormatter, but after explicitly setting the tick labels, it becomes a FixedFormatter:

>>> f, ax = plt.subplots()
>>> import numpy as np
>>> plt.plot(np.arange(100))
[<matplotlib.lines.Line2D object at 0x2cee250>]
>>> ax.xaxis.get_major_formatter()
<matplotlib.ticker.ScalarFormatter object at 0x2cdf490>
>>> ax.set_xticklabels(['foo %s' % i for i in range(100)])
[<matplotlib.text.Text object at 0x2cd9990>, <matplotlib.text.Text object at 0x2cdf5d0>, <matplotlib.text.Text object at 0x2ceee10>, <matplotlib.text.Text object at 0x2cf3590>, <matplotlib.text.Text object at 0x2cf3cd0>, <matplotlib.text.Text object at 0x2cf6450>]
>>> ax.xaxis.get_major_formatter()
<matplotlib.ticker.FixedFormatter object at 0x2cee410>
>>> 

[WeatherGod]

FixedFormatter.format_data_short() falls back to Formatter.format_data_short(), which utilizes __call__(). For a FixedFormatter, __call__() is defined as:

    def __call__(self, x, pos=None):
        'Return the format for tick val *x* at position *pos*'
        if pos is None or pos >= len(self.seq):
            return ''
        else:
            return self.seq[pos]

So, the position is assumed to be an index into the sequence, but the data coordinate system isn't necessarially set up that way, I think.

[WeatherGod]

ah, it is even easier than that! pos is None!

[WeatherGod]

syntax error. Missing parens.

[tacaswell]

This breaks the use case where you do

from matplotlib.ticker import FixedLocator, FixedFormatter

fig, ax = plt.subplots()
ax.xaxis.set_major_locator(FixedLocator([10, 20, 30]))
ax.xaxis.set_major_formatter(FixedFormatter(['a', 'b', 'c']))
ax.set_xlim([5, 35])

[WeatherGod]

ah, and I took a look at your changes and I see you arrived at the same spot in the code in a different way. So, what was "pos" ever for?

[tacaswell]

I came across this a while ago and thought it was a feature not a bug ;). If you are using a fixed formatter, then you have de-coupled the tick labels from the data so there is no sensible way we can guess at what the values between the ticks are.

I think pos is passed is as part of the call to the formatter from the Axis code.

[tacaswell]

See https://github.com/matplotlib/matplotlib/blob/master/lib/matplotlib/axis.py#L873 for where the formatter is called as part of the Axis draw process (which led me down the rabbit hole of how Axis objects draw them selves.... They seem to have a pool of Tick objects and then update and draw just enough of them. The Tick objects wrap up both tick labels, the tick mark, and the grid line.).

[mdboom]

Ah, of course. If the labels are not 1-to-1 with values this isn't going to work. Thanks.