BUG: Fix nanvar for large float16 arrays #40738
Conversation
Summary of bug: Above a certain length, calling `.std() and `.var()` on a Series or DataFrame fails for float types. The is easily shown with float16 in the following example: ```python >>> import pandas as pd >>> import numpy as np >>> zeros = 32760 >>> ones = 32759 >>> pd.Series([0] * zeros + [1] * ones).astype("float16").var() 0.25 >>> zeros = 32760 >>> ones = 32760 >>> pd.Series([0] * zeros + [1] * ones).astype("float16").var() 0.0 ``` What is happening here is that both `count` and `d` overflow to `inf` as 65519 is the largest integer float16 can represent. Example: ```python >>> import numpy as np >>> np.float16(65519) 65500.0 >>> np.float16(65520) inf ``` This makes the variance go to 0.0. Since the `avg` calculation on 927 casts to float64 anyway, it seems like there is no downside to not casting `count` and `d`. It will also improve accuracy in cases where the overflow does not happen. INSTALLED VERSIONS ------------------ commit : f2c8480 python : 3.8.2.final.0 python-bits : 64 OS : Darwin OS-release : 20.3.0 Version : Darwin Kernel Version 20.3.0: Thu Jan 21 00:07:06 PST 2021; root:xnu-7195.81.3~1/RELEASE_X86_64 machine : x86_64 processor : i386 byteorder : little LC_ALL : None LANG : en_GB.UTF-8 LOCALE : en_GB.UTF-8 pandas : 1.2.3 numpy : 1.20.1 | Seems like the unit test failures are related to matplotlib not this change so I do not think this logic is tested. If someone agrees this is a useful fix, I am happy to make an issue, add tests and make a whatsnew entry. |
jreback left a comment
jreback left a comment There was a problem hiding this comment.
we have almost 0 float16 support, what exactly are you trying to fix?
@jreback I do agree with you float16 is very niche, but since this isn't a huge change and is something that is support in numpy (see below) I think it's worth changing. The problem is that once you have a DataFrame/Series longer than 65519 entries, it will count those entries and cast it to a float16. This will fail as it can't be represented so will be >>> import numpy as np >>> zeros = 32760 >>> ones = 32760 >>> np.array([0] * zeros + [1] * ones).astype("float16").var() 0.2498It is also not an issue with |
| To clarify the bug in one line: Any Series or DataFrame with type float16 that has more than 65519 entries will unexpectedly return 0.0 variance and standard deviation, no matter their values and not in accordance with numpy. |
jreback added Algos | if mask is not None: | ||
| values[mask] = np.nan | ||
| | ||
| if is_float_dtype(values.dtype): |
There was a problem hiding this comment.
if it doesn't break anything else would take the change, pls add a test and a whatsnew note
| This pull request is stale because it has been open for thirty days with no activity. Please update or respond to this comment if you're still interested in working on this. |
| closing this as stale |
| ping if you would like to address and can reopen |
2 participants
Summary of bug:
Above a certain length, calling
.std()and.var()on a Series or DataFrame fails for float types. The is easily shown with float16 in the following example:What is happening here is that both
countanddoverflow toinfas 65519 is the largest integer float16 can represent. Example:This makes the variance go to 0.0. Since the
avgcalculation on line 927 of this file casts to float64 anyway, it seems like there is no downside to not castingcountandd. It will also improve accuracy in cases where the overflow does not happen.INSTALLED VERSIONS
commit : f2c8480
python : 3.8.2.final.0
python-bits : 64
OS : Darwin
OS-release : 20.3.0
Version : Darwin Kernel Version 20.3.0: Thu Jan 21 00:07:06 PST 2021; root:xnu-7195.81.3~1/RELEASE_X86_64
machine : x86_64
processor : i386
byteorder : little
LC_ALL : None
LANG : en_GB.UTF-8
LOCALE : en_GB.UTF-8
pandas : 1.2.3
numpy : 1.20.1