Repository navigation
Closed
Description
Bug report
Bug summary
The following codes fail in matplotlib 3.1.1, while work properly in 3.0.3
case 1
Code for reproduction
import pandas as pd
from matplotlib import pyplot as plt
df = pd.DataFrame({"x":[1,2,3],"width":[.2,.4,.6]},index=[1,2,3])
plt.figure()
plt.bar(df.x, 1, width=df.width)
plt.show()Actual outcome
Traceback (most recent call last):
File "/Desktop/fail_example.py", line 7, in <module>
plt.bar(df.x, 1, width=df.width)
File "/.pyenv/versions/3.7.3/lib/python3.7/site-packages/matplotlib/pyplot.py", line 2440, in bar
**({"data": data} if data is not None else {}), **kwargs)
File "/.pyenv/versions/3.7.3/lib/python3.7/site-packages/matplotlib/__init__.py", line 1601, in inner
return func(ax, *map(sanitize_sequence, args), **kwargs)
File "/.pyenv/versions/3.7.3/lib/python3.7/site-packages/matplotlib/axes/_axes.py", line 2430, in bar
label='_nolegend_',
File "/.pyenv/versions/3.7.3/lib/python3.7/site-packages/matplotlib/patches.py", line 707, in __init__
Patch.__init__(self, **kwargs)
File "/.pyenv/versions/3.7.3/lib/python3.7/site-packages/matplotlib/patches.py", line 89, in __init__
self.set_linewidth(linewidth)
File "/.pyenv/versions/3.7.3/lib/python3.7/site-packages/matplotlib/patches.py", line 368, in set_linewidth
self._linewidth = float(w)
Expected outcome
Outcome in 3.0.3:

case 2
Code for reproduction
import pandas as pd
from matplotlib import pyplot as plt
df = pd.DataFrame({"x":[1,3,10]},index=[1000,2000,3000])
plt.figure()
plt.bar(df.x, 1, width=.2)
plt.show()Expected outcome
Outcome in 3.0.3:

Matplotlib version
- Operating system: MacOSX
- Matplotlib version: 3.1.1 (installed from pip)
- Matplotlib backend (
print(matplotlib.get_backend())): MacOSX - Python version: 3.7.3
- Jupyter version (if applicable):
- Other libraries: pandas 0.24.2
Cause of the bug
_axes.py, line 2363 in bar:
x0 = x
x = np.asarray(self.convert_xunits(x))
width = self._convert_dx(width, x0, x, self.convert_xunits)
_axes.py, line 2166 in _convert_dx:
try:
x0 = x0[0]
except (TypeError, IndexError, KeyError):
x0 = x0
I guess x0 is expected to be a scalar after this line, However when x0 is an int-indexed (and 0 is not in its indices) pandas.Series, it is continued to be pandas.Series and causes chained disorders to end in the error above.
To evidence, they worked properly when rewritten as (ugly!):
try:
x0 = x0[0]
except (TypeError, IndexError, KeyError):
try:
x0 = x0.iat[0]
except:
x0 = x0

Activity
[-]Axes._convert_dx(dx,x0,x,convert) fails when x0 is an int-indexed pandas.Series[/-][+]bar(Axes._convert_dx(dx,x0,x,convert) fails when x0 is an int-indexed pandas.Series[/+][-]bar(Axes._convert_dx(dx,x0,x,convert) fails when x0 is an int-indexed pandas.Series[/-][+]axes.bar fails when x and width are int-indexed pandas.Series[/+][-]axes.bar fails when x and width are int-indexed pandas.Series[/-][+]axes.bar fails when x is int-indexed pandas.Series[/+]ImportanceOfBeingErnest commentedon Aug 31, 2019
This came in with #12903.
Worth noting that
plt.bar(df.x, 1, width=.1)also produces wrong output (though there is no error in that case).jklymak commentedon Aug 31, 2019
Sorry for the bug. I guess we could just put the old messy code back in, but it'd be nice if someone who understood pandas data types and inheritance figured out the proper solution. Obviously if you remove the index argument this works, or changing the index to zero-indexing. Is
iatthe way to get the first element in the series regardless of how its been indexed? Its inelegant of us to have special pandas magic all over our code, so it would be nice if something turned x into a proper numpy array earlier.ImportanceOfBeingErnest commentedon Aug 31, 2019
Apparently
.iatis preferred over.iloc.jklymak commentedon Aug 31, 2019
Yeah but why doesn’t df.x[0] return the first element?
ImportanceOfBeingErnest commentedon Aug 31, 2019
For the same reason
{1 : True}[0]raises an error I suppose. The confusing bit is just thatdf.index[0]actually works, because an index behaves differently than a series.jklymak commentedon Aug 31, 2019
OK, well, this comes back to the units parsing. We can't just
np.asarraybecause the way some packages implement units, this kills the units information, and we have no good way to get the units out to save for later. But, if we don't do this, then pandas doesn't work if someone manually sets the indices becausex[0]no longer works. Yes, we could keep special casing every possible way data gets passed to functions, or we could come up with a few rules that we always apply to everything that comes in.The obvious stopgap is to just put in the
x.iat[0], but again, I really think that its bad engineering to have triple-nested try/except blocks in the code as we try and guess what kind of data we have been passed. So I'll keep this open to hopefully engender discussion.ImportanceOfBeingErnest commentedon Aug 31, 2019
Is there any alternative to finding the
typeof the first element? The culprit seems to be the unit package that subclasses numpy.ndarray, right? So unless that package has some other means to know what the array contains, one would have no choice but to stick to the currently solution. And in that case, special-casing pandas series is the only way out. Instead of try..except one could also usehasattr(x, "iat"), or"series" in str(type(x)).split["."]or so.jklymak commentedon Aug 31, 2019
I think I have a solution that squares all the constraints, but avoids using a library-specific method like
iat. I’ll post later tonight.tacaswell commentedon Sep 1, 2019
It is better to think of data frames as "dicts of Series with a bunch of helper methods" and Series as "ordered dicts of values with an optional index based lookup and helper methods".
I have lobbied Jeff Reback to get the behavior of
[]change to be positional (aka integer) by default, but a) that would cause an insane amount of code breakage b) there is a big constituency of pandas users who have non-trivial indexes and benefit froms.loc[k]/s.at[k](https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.loc.html the "label" based one) being fast-pathed rather ands.iloc[k]/s.iat[k]. The extra tricky thing is if your labels happen to be integers it looks like they behave the same. Everyone has questionable early design decision that they can not back out of 🤷♀️ .