idxmax() not available on groupby anymore

[jorisvandenbossche]

Just saw an example in this SO question, the use of idxmax() on a groupby object:

df.groupby(...).idxmax()

This worked in 0.12, but not anymore in 0.13 as it is not in the whitelist.

• Should this be added to the whitelist? (idxmax/idxmin for SeriesGroupby) I think this is a usefull method on a groupby object.
• I got an AttributeError and not a DeprecationWarning. There was spoken of a deprecation first (BUG: value_counts / shift not in groupby dispatch whitelist (0.12-dev) #5480), but was it decided not to do it?

[jreback]

you can add to whitelist
no depreciation warning anymore because u also need to maintain a blacklist which makes maintaining this even harder

FYI

you can always do

df.groupby(...).apply(lambda x: x.idxmax())

[jorisvandenbossche]

@jreback I know, but for idxmax I think this is usefull enough to have to shortcut available. Plus, this will break existing code (as the example on SO shows that it is used).

And I was just about to open an new issue about this, because it seems the apply also does not work the same as in 0.12. With example from this SO (http://stackoverflow.com/questions/20819702/hourly-frequency-count-with-python/20820166#20820166), and with 0.13rc, this:

df.groupby(df.index.date).apply(lambda x: x.idxmax())

generates a

ValueError: Shape of passed values is (1, 3), indices imply (1, 3)

while in 0.12 it does work.

It has to do I think with difference between SeriesGroupby and DataFrameGroupby with one column, as df.groupby(df.index.date)['value'].apply(lambda x: x.idxmax()) does work in master, but in 0.12 both ways work.

[jreback]

ok go ahead and do a or for adding idxmax/min to whitelist (and in the test for same)

I'll look at the apply issue

[jreback]

can u open the apply as a separate issue with a small example
thanks (so close this one with a pr for adding idxmax/min to whitelist)

[jorisvandenbossche]

@jreback opened new issue #5788

[jreback]

thanks for pr and issues!