BUG: fix fill value for gouped sum in case of unobserved categories for string dtype (empty string instead of 0) #61909
Conversation
…or string dtype (empty string instead of 0)
jorisvandenbossche
added
Bug
Groupby
Strings
pandas/_libs/groupby.pyx
Outdated
| @@ -729,6 +730,10 @@ def group_sum( | |||
| sumx = np.zeros((<object>out).shape, dtype=(<object>out).base.dtype) | |||
| compensation = np.zeros((<object>out).shape, dtype=(<object>out).base.dtype) | |||
|
|
|||
| if is_string: | |||
| # for strings start with empty string instead of 0 as `initial` value | |||
| sumx = np.full((<object>out).shape, "", dtype=object) | |||
There was a problem hiding this comment.
Would passing “initial” be more general/idiomatic?
There was a problem hiding this comment.
I was initially thinking that as well, but then because this would in practice only be used for the specific case of strings, I thought to be more explicit about that fact. But in both cases I have to pass it down a few layers, so whether it is then called initial or is_string actually does not matter much, and initial at least makes it clearer where it is called from the EA _groupby_ops what the purpose is of passing it down.
So updated to use initial in the last commit
jorisvandenbossche
deleted the
string-dtype-groupby-sum-observed-false-fillvalue
branch
|
Owee, I'm MrMeeseeks, Look at me. There seem to be a conflict, please backport manually. Here are approximate instructions:
And apply the correct labels and milestones. Congratulations — you did some good work! Hopefully your backport PR will be tested by the continuous integration and merged soon! Remember to remove the If these instructions are inaccurate, feel free to suggest an improvement. |
…or string dtype (empty string instead of 0) (pandas-dev#61909)
|
Manual backport -> #61963 |
I ran into one more case of the sum of empty / all-NaN to use "0" instead of empty string (#60229), specifically when effectively introducing empty groups with categorical data with observed=False.
Follow-up on #60936
The passing through of
is_stringthrough several layers is a bit annoying, but effectively is needed to for now only changes this for string dtype, and not for object dtype in general (which in the other PR related to this, we did for now)