WIP/BUG: Correct results for groupby(...).transform with null keys #45839
Conversation
rhshadrach added Bug Groupby Apply 
pandas/core/groupby/groupby.py Outdated
| # return [self.indices.get(name, []) for name in names] | ||
| # self.indices is a dict and doesn't handle looking up nulls in the groups | ||
| from pandas import ( | ||
| Index, | ||
| Series, | ||
| ) | ||
| | ||
| index = Index(self.indices.keys(), tupleize_cols=False) | ||
| indices = Series(self.indices.values(), index=index) | ||
| result = [indices[name] for name in names] | ||
| return result |
There was a problem hiding this comment.
Here self.indices is a dict mapping the name to the indices it occurs at. Problem is when there are null values for the name (e.g. np.nan, (1, np.nan), etc). Here the lookup fails, and in general one shouldn't have null values as keys in a dict. I know Series gets around this using codes, hence the solution above.
The need to separate out the Index carefully is to handle odd cases like where self.indices is {(1,): np.array([0, 2]), (1, 2): np.array([1, 3])}. This is the particular case for a single test.
cc @jbrockmendel @phofl if there are any ideas to solve this in some more direct way rather than using a Series.
There was a problem hiding this comment.
Here the lookup fails, and in general one shouldn't have null values as keys in a dict
Not sure if this is relevant, but possibly related to #43943?
There was a problem hiding this comment.
i think in index.pyx we have some code for canonicalizing NA values, e.g. float("nan") -> np.nan, so that dict lookups are better-behaved. could be adapted/reused when defining indices?
There was a problem hiding this comment.
Thanks - this is precisely what I was looking for. However, I found a simpler way to use the codes directly; I've put up #45953.
pandas/core/groupby/groupby.py Outdated
| Series, | ||
| ) | ||
| | ||
| index = Index(self.indices.keys(), tupleize_cols=False) |
There was a problem hiding this comment.
can you add a comment about why this is two lines instead of just indices = Series(self.indices)? i assume its to avoid a MultiIndex
2 participants
transformgroupbywithNaNs #17093 (Replace xxxx with the Github issue number)doc/source/whatsnew/vX.X.X.rstfile if fixing a bug or adding a new feature.Fixes a lot of issues with
groupby(...).transformwhen there are null values in the groups. Still needs tests for SeriesGroupBy. Happy to break this up if requested, but I'm stuck on one part; see comment below.Note this changes some tested behavior, but it changes it to conform to what is specified in the docs from DataFrameGroupBy.transform