I just updated pandas from 0.17.1 to 0.21.0 to take advantage of some new functionalities, and ran into compatibility issue with matplotlib (which I also updated to latest 2.1.0). In particular, the Timestamp object seems to be changed significantly.
I happen to have another machine still running the older versions of pandas(0.17.1)/matplotlib(1.5.1) which I used to compared the differences:
Both versions show my DataFrame index to be dtype='datetime64[ns]
DatetimeIndex(['2017-03-13', '2017-03-14', ... '2017-11-17'], type='datetime64[ns]', name='dates', length=170, freq=None)
But when calling type(df.index[0])
, 0.17.1 gives pandas.tslib.Timestamp
and 0.21.0 gives pandas._libs.tslib.Timestamp
.
When plotting with df.index
as x-axis:
plt.plot(df.index, df['data'])
matplotlibs by default formats the x-axis labels as dates for pandas 0.17.1 but fails to recognize it for pandas 0.21.0 and simply gives raw number 1.5e18
(epoch time in nanosec).
I also have a customized cursor that reports clicked location on the graph by using matplotlib.dates.DateFormatter
on the x-value which fails for 0.21.0 with:
OverflowError: signed integer is greater than maximum
I can see in debug the reported x-value is around 736500 (i.e. day count since year 0) for 0.17.1 but is around 1.5e18 (i.e. nanosec epoch time) for 0.21.0.
I am surprised at this break of compatibility between matplotlib and pandas as they are obviously used together by most people. Am I missing something in the way I call the plot function above for the newer versions?
Update as I mentioned above, I prefer directly calling plot
with a given axes object but just for the heck of it, I tried calling the plot method of the DataFrame itself df.plot()
. As soon as this is done, all subsequent plots correctly recognize the Timestamp within the same python session. It's as if an environment variable is set, because I can reload another DataFrame or create another axes with subplots
and no where does the 1.5e18
show up. This really smells like a bug as the latest pandas doc says pandas:
The plot method on Series and DataFrame is just a simple wrapper around plt.plot()
But clearly it does something to the python session such that subsequent plots deal with the Timestamp index properly.
In fact, simply running the example at the above pandas link:
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
ts = pd.Series(np.random.randn(1000), index=pd.date_range('1/1/2000', periods=1000))
Depending on whether ts.plot()
is called or not, the following plot either correctly formats x-axis as dates or not:
plt.plot(ts.index,ts)
plt.show()
Once a member plot is called, subsequently calling plt.plot
on new Series or DataFrame will autoformat correctly without needing to call the member plot method again.
There is an issue with pandas datetimes and matplotlib coming from the recent release of pandas 0.21, which does not register its converters any more at import. Once you use those converters once (within pandas) they'll be registered and automatically used by matplotlib as well.
A workaround would be to register them manually,
import pandas.plotting._converter as pandacnv
pandacnv.register()
In any case the issue is well known at both pandas and matplotlib side, so there will be some kind of fix for the next releases. Pandas is thinking about readding the register in an upcomming release. So this issue may be there only temporarily. An option is also to revert to pandas 0.20.x where this should not occur.
Update: this is no longer an issue with current versions of matplotlib (2.2.2)/pandas(0.23.1), and likely many that have been released since roughly December 2017, when this was fixed.
Update 2: As of pandas 0.24 or higher the recommended way to register the converters is
from pandas.plotting import register_matplotlib_converters
register_matplotlib_converters()
or if pandas
is already imported as pd
,
pd.plotting.register_matplotlib_converters()
After opening an issue on pandas github, I learned that this was indeed a known issue between pandas and matplotlib regarding auto-registration of unit converter. In fact it was listed on what's new page which I had failed to see before, along with the proper way to register the converters:
from pandas.tseries import converter
converter.register()
This is also done the first time a member plot method is called on a Series or DataFrame which explains what I observed above.
It appears to have been done with the intention that matplotlib is supposed to implement some basic support for pandas datetime, but indeed a deprecation warning of some sort could be useful for such a break. However until matplotlib actually implements such support (or some sort of lazy registration mechanism), practically I'm always putting those two lines at the pandas import. So I'm not sure why pandas would want to disable the automatic registration on import before things are ready on the matplotlib side.
It looks this issue has been fixed in the future version of matplotlib.
Try to run "pip install --upgrade matplotlib".
I met the same issue "AttributeError: 'numpy.datetime64' object has no attribute 'toordinal'". It was fixed when upgrade the package matplotlib.
If you love us? You can donate to us via Paypal or buy me a coffee so we can maintain and grow! Thank you!
Donate Us With