argmin/argmax, idxmin/idxmax and datetime64

[lodagro]

In [65]: s
Out[65]: 
a   2013-01-02 00:00:00
b   2013-01-03 00:00:00
c   2013-01-04 00:00:00
d   2013-01-05 00:00:00
e                   NaT
dtype: datetime64[ns]

In [66]: s.argmin()
Out[66]: 4

In [67]: s.idxmin()
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
...
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'

In [68]: s.argmax()
Out[68]: 3

In [69]: s.idxmax()
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
...
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'

[jreback]

yep....I had change a few things...

[jreback]

thanks for the bug report...fixed in the pr above (and it fixes the failing test on 32-bit in argsort)

....in your example, s.argmin() is WRONG (because its calling the numpy argmin), argmax works (because NaT is a really big negative int). Not sure if should open an issue on this. (would essentially have them do self.index.get_loc(self.idxmin()). worth adding this feature?

[jreback]

I think u can close this

and if u want to open an issue about argmin/max

[lodagro]

idxmax/idxmin are fine now! So ill close this one, as you proposed. Concerning argmin/argmax, ill leave it up to you to decide if those needs to be adressed/overloaded.

In [76]: s
Out[76]: 
a   2013-01-02 00:00:00
b   2013-01-03 00:00:00
c   2013-01-04 00:00:00
d   2013-01-05 00:00:00
e                   NaT
dtype: datetime64[ns]

In [77]: s.idxmin()
Out[77]: 'a'

In [78]: s.idxmax()
Out[78]: 'd'

Thanks!

[lrq3000]

I'm getting the same issue with Pandas v0.11 stable on a DataFrame with dtype object (but it only contains float64 values), and also argmax() seems to have been deleted, it's not available anymore for DafaFrame.

/EDIT: Just casted the DataFrame into a float64 and it works again. But it's weird that idxmax() fails with object dtype with the error above, maybe another conditional check may be added to avoid failing like this?

[jreback]

you cannot use dtype of object (well you can), but it won't work well with datetimelike, use datetime64[ns], pls post what you are doing and i'll help you out

[lrq3000]

Thank you jreback but it's ok, my DataFrame is in fact only composed of floats (or strings which are only composed of numbers but then it's ok) and thus recasting into floats solved the problem, but it was weird because Pandas automatically casted my DataFrame as an object when there are only floats inside, that's why I was a bit surprised.

[jreback]

@lrq3000 glad that it worked, but as a general rule, object type is not very efficicient, that is why its very importatnt to use native dtypes.

[gerritholl]

Related: numpy/numpy#5610