Pandas version-specific error: TypeError: ufunc 'isfinite' not supported for the input types

[jaradc]

With pandas version 0.20.3, the below code works. With pandas version 0.21.0, it throws the following error:

Traceback (most recent call last):
  File "C:/Users/jarad/Downloads/ffdsfee33.py", line 13, in <module>
    model.plot(fc)
  File "C:\Python35\lib\site-packages\fbprophet\forecaster.py", line 1260, in plot
    alpha=0.2)
  File "C:\Python35\lib\site-packages\matplotlib\__init__.py", line 1889, in inner
    return func(ax, *args, **kwargs)
  File "C:\Python35\lib\site-packages\matplotlib\axes\_axes.py", line 4770, in fill_between
    x = ma.masked_invalid(self.convert_xunits(x))
  File "C:\Python35\lib\site-packages\numpy\ma\core.py", line 2388, in masked_invalid
    condition = ~(np.isfinite(a))
TypeError: ufunc 'isfinite' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''

Re-creatable code example:

df = pd.read_clipboard(parse_dates=['ds'])

model = Prophet(yearly_seasonality=False, daily_seasonality=False)
model.fit(df)
future_df = model.make_future_dataframe(30)
fc = model.predict(future_df)
model.plot(fc)
plt.show()

ds	y
5/15/2017	0.3664000
5/16/2017	0.3728000
5/17/2017	0.3596000
5/18/2017	0.3633399
5/19/2017	0.3576137
5/20/2017	0.3332930
5/21/2017	0.3808267
5/22/2017	0.3673000
5/23/2017	0.3587394
5/24/2017	0.3329000
5/25/2017	0.3492000
5/26/2017	0.3566112
5/27/2017	0.3615000
5/28/2017	0.3814000
5/29/2017	0.3465107
5/30/2017	0.3516000
5/31/2017	0.3298000
6/1/2017	0.3623000
6/2/2017	0.3609819
6/3/2017	0.3652000
6/4/2017	0.3728000
6/5/2017	0.3399985
6/6/2017	0.3586000
6/7/2017	0.3574000
6/8/2017	0.3727000
6/9/2017	0.3732567
6/10/2017	0.3725000
6/11/2017	0.3743000
6/12/2017	0.3495000
6/13/2017	0.3641000
6/14/2017	0.3585000
6/15/2017	0.3421000
6/16/2017	0.3513149
6/17/2017	0.3568000
6/18/2017	0.3457000
6/19/2017	0.3530000
6/20/2017	0.3397578
6/21/2017	0.3333000
6/22/2017	0.3527577
6/23/2017	0.3689000
6/24/2017	0.3743599
6/25/2017	0.3785000
6/26/2017	0.3629000
6/27/2017	0.3451601
6/28/2017	0.3674000
6/29/2017	0.3550000
6/30/2017	0.3590000
7/1/2017	0.3846000
7/2/2017	0.3628000
7/3/2017	0.3667000
7/4/2017	0.3711000
7/5/2017	0.3421000
7/6/2017	0.3645967
7/7/2017	0.3794752
7/8/2017	0.3929000
7/9/2017	0.3857000
7/10/2017	0.3760494
7/11/2017	0.3441000
7/12/2017	0.3666322
7/13/2017	0.3584000
7/14/2017	0.3691000
7/15/2017	0.3890000
7/16/2017	0.3824000
7/17/2017	0.3903000
7/18/2017	0.3726750
7/19/2017	0.3433075
7/20/2017	0.3424415
7/21/2017	0.3484333
7/22/2017	0.3838619
7/23/2017	0.3607334
7/24/2017	0.3624161
7/25/2017	0.3445446
7/26/2017	0.3652995
7/27/2017	0.3514286
7/28/2017	0.3648852
7/29/2017	0.3706496
7/30/2017	0.3572222
7/31/2017	0.3588297
8/1/2017	0.3374567
8/2/2017	0.3488459
8/3/2017	0.3591218
8/4/2017	0.3396043
8/5/2017	0.3636433
8/6/2017	0.3542834
8/7/2017	0.3449915
8/8/2017	0.3338922
8/9/2017	0.3516176
8/10/2017	0.3404888
8/11/2017	0.3492653
8/12/2017	0.3771050
8/13/2017	0.3621665
8/14/2017	0.3459562
8/15/2017	0.3319234
8/16/2017	0.3364871
8/17/2017	0.3464671
8/18/2017	0.3455390
8/19/2017	0.3291302
8/20/2017	0.3208030
8/21/2017	0.3321821
8/22/2017	0.3337696
8/23/2017	0.3289861
8/24/2017	0.3526127
8/25/2017	0.3611672
8/26/2017	0.3746895
8/27/2017	0.3720981
8/28/2017	0.3359440
8/29/2017	0.3726067
8/30/2017	0.3380052
8/31/2017	0.3473166
9/1/2017	0.3698962
9/2/2017	0.3783205
9/3/2017	0.3753667
9/4/2017	0.3897302
9/5/2017	0.3608144
9/6/2017	0.3454317
9/7/2017	0.3704847
9/8/2017	0.3772438
9/9/2017	0.3844822
9/10/2017	0.3872648
9/11/2017	0.3544398
9/12/2017	0.3474683
9/13/2017	0.3623862
9/14/2017	0.3415039
9/15/2017	0.3512003
9/16/2017	0.3580232
9/17/2017	0.3669942
9/18/2017	0.3522143
9/19/2017	0.3450433
9/20/2017	0.3420750
9/21/2017	0.3497456
9/22/2017	0.3705893
9/23/2017	0.3822981
9/24/2017	0.3886189
9/25/2017	0.3670670
9/26/2017	0.3592365
9/27/2017	0.3236036
9/28/2017	0.3651484
9/29/2017	0.3494884
9/30/2017	0.3788507
10/1/2017	0.3681230
10/2/2017	0.3402499
10/3/2017	0.3653453
10/4/2017	0.3616628
10/5/2017	0.3300067
10/6/2017	0.3489849
10/7/2017	0.3541858
10/8/2017	0.3404519
10/9/2017	0.3457787
10/10/2017	0.3483314
10/11/2017	0.3521676
10/12/2017	0.3784831
10/13/2017	0.3974615
10/14/2017	0.3769376
10/15/2017	0.3982707
10/16/2017	0.3665458
10/17/2017	0.4034299
10/18/2017	0.4028199
10/19/2017	0.3946326
10/20/2017	0.4132913
10/21/2017	0.4357214
10/22/2017	0.4370657
10/23/2017	0.4035825
10/24/2017	0.4017471
10/25/2017	0.3896750
10/26/2017	0.4086555
10/27/2017	0.3748427
10/28/2017	0.3646382
10/29/2017	0.4090890

[kilianbreathnach]

I have found this problem too. It seems to be an issue with numpy:

numpy/numpy#5610

[bletham]

It was just pointed out in #33 that this issue can come from the way matplotlib tries to use pandas datetimes. @JKPoynter also provides a fix, I'll get it in soon. In the meantime, you can pass uncertainty=False to the plot method and it will work just without uncertainty intervals. You could then add fcst['yhat_lower'] and fcst['yhat_upper'] to the figure using the fix from #33.

[bletham]

This is fixed in d3d1fcd by using python datetime instead of numpy for the plots. If you could install from github master branch and check if things are working again for you that would be helpful.

[jaradc]

I just tested and at first it didn't work for me. What got it to work for me was updating my matplotlib package. Version '2.0.0rc2' was installed. I updated to '2.1.0' and that solved it.

For verbosity, I:
updated pystan
installed fbprophet from github python setup.py (not pip install fbprophet)
installed pandas version 0.21.0 (to create same scenario in my original question)

those steps didn't seem to work.

I then update matplotlib from '2.0.0rc2' to '2.1.0'. I hope all this wasn't because I was using a development version of matplotlib. :( If so, I am so sorry! I hope this helps or adds some value in some way.

Synopsis: everything seems to be working as-expected now.

[bletham]

I can replicate the issue with PyPI fbprophet, pandas 0.21 and matplotlib 2.0.0, so it is not a matter of being a release candidate.

I can also confirm that upgrading to matplotlib 2.1.0 fixes the issue with PyPI fbprophet, so it seems to be a combination of pandas >= 0.21.0 and matplotlib < 2.1.0 that is the problem.

For me, however, the current github master branch fbprophet works with pandas 0.21.0 and matplotlib 2.0.0. I'd prefer this fix to the issue as opposed to setting a matplotlib version requirement of 2.1.0. I'm hesitant to put the matplotlib requirement since people may have a lot of other things depending on matplotlib and the upgrade could be a pain. I know you said that for you the issue persisted after installing the github version. I'm wondering if maybe the PyPI version didn't uninstall completely or something like that. If it isn't too big a pain to downgrade matplotlib to recreate this issue (or if anyone else can currently produce this issue), would you mind just making sure that fbprophet is uninstalled (can't be imported), install the github master branch, and then post the full stacktrace of the error so I can see what line in Prophet.plot is raising the error?

[jaradc]

1. Try to import fbprophet, or matplotlib. Both don't import.

>>> import fbprophet
Traceback (most recent call last):
  File "<pyshell#0>", line 1, in <module>
    import fbprophet
ImportError: No module named 'fbprophet'
>>> import matplotlib
Traceback (most recent call last):
  File "<pyshell#1>", line 1, in <module>
    import matplotlib
ImportError: No module named 'matplotlib'

2. Install matplotlib version 2.0.0rc2:

C:\Windows\System32>pip install matplotlib==2.0.0rc2
Collecting matplotlib==2.0.0rc2
  Downloading matplotlib-2.0.0rc2-cp35-cp35m-win_amd64.whl (50.7MB)
    100% |################################| 50.7MB 21kB/s
Requirement already satisfied: numpy>=1.7.1 in c:\python35\lib\site-packages (fr
om matplotlib==2.0.0rc2)
Requirement already satisfied: six>=1.10 in c:\python35\lib\site-packages (from
matplotlib==2.0.0rc2)
Requirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=1.5.6 in c:\py
thon35\lib\site-packages (from matplotlib==2.0.0rc2)
Requirement already satisfied: cycler>=0.10 in c:\python35\lib\site-packages (fr
om matplotlib==2.0.0rc2)
Requirement already satisfied: python-dateutil in c:\python35\lib\site-packages
(from matplotlib==2.0.0rc2)
Requirement already satisfied: pytz in c:\python35\lib\site-packages (from matpl
otlib==2.0.0rc2)
Installing collected packages: matplotlib
Successfully installed matplotlib-2.0.0rc2

3. Install the github master branch version of fbprophet.
   When in /prophet/python, I just did python setup.py install

4. Check imports and versions to verify (interactive):

>>> import matplotlib
>>> import fbprophet
>>> matplotlib.__version__
'2.0.0rc2'
>>> fbprophet.__version__
'0.2.dev'

5. Execute example:

>>> import pandas as pd
>>> df = pd.read_clipboard(parse_dates=['ds'])
>>> from fbprophet import Prophet
>>> import matplotlib.pyplot as plt
>>> model = Prophet(yearly_seasonality=False, daily_seasonality=False)
>>> model.fit(df)
<fbprophet.forecaster.Prophet object at 0x0000000008DEBF98>
>>> future_df = model.make_future_dataframe(30)
>>> fc = model.predict(future_df)
>>> model.plot(fc)
<matplotlib.figure.Figure object at 0x000000000387C240>
>>> plt.show()

plot showed as expected with matplotlib version 2.0.0rc2.

If I uninstall fbprophet 0.2.dev and install fbprophet from pip install fbprophet (version 0.2), I get the error again.

>>> from fbprophet import Prophet
>>> import pandas as pd
>>> df = pd.read_clipboard(parse_dates=['ds'])
>>> model = Prophet(yearly_seasonality=False, daily_seasonality=False)
>>> model.fit(df)
<fbprophet.forecaster.Prophet object at 0x00000000034C6860>
>>> future_df = model.make_future_dataframe(30)
>>> fc = model.predict(future_df)
>>> model.plot(fc)
Traceback (most recent call last):
  File "<pyshell#7>", line 1, in <module>
    model.plot(fc)
  File "C:\Python35\lib\site-packages\fbprophet\forecaster.py", line 1260, in plot
    alpha=0.2)
  File "C:\Python35\lib\site-packages\matplotlib\__init__.py", line 1889, in inner
    return func(ax, *args, **kwargs)
  File "C:\Python35\lib\site-packages\matplotlib\axes\_axes.py", line 4770, in fill_between
    x = ma.masked_invalid(self.convert_xunits(x))
  File "C:\Python35\lib\site-packages\numpy\ma\core.py", line 2388, in masked_invalid
    condition = ~(np.isfinite(a))
TypeError: ufunc 'isfinite' not supported for the input types, and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''
>>> fbprophet.__version__
Traceback (most recent call last):
  File "<pyshell#8>", line 1, in <module>
    fbprophet.__version__
NameError: name 'fbprophet' is not defined
>>> import fbprophet
>>> fbprophet.__version__
'0.2'

So my understanding from all this is that the fbprophet 0.2.dev version fixed the problem. And, I must have not fully uninstalled everything correctly in my last comment.

[bletham]

@jaradc that's great, thanks a bunch for your help in debugging this!

[bletham]

The fix has been pushed to PyPI in v0.2.1.