BUG: DTI.value_counts doesnt preserve tz

[sinhrks]

Found 2 problems related to value_counts.

• DatetimeIndex.value_counts loses tz.

didx = pd.date_range('2011-01-01 09:00', freq='H', periods=3, tz='Asia/Tokyo')
print(didx.value_counts())
#2011-01-01 00:00:00    1
#2011-01-01 01:00:00    1
#2011-01-01 02:00:00    1
# dtype: int64

• PeriodIndex.value_counts results in Int64Index, and unable to drop NaT.

pidx = pd.PeriodIndex(['2011-01-01 09:00', '2011-01-01 10:00', pd.NaT], freq='H')
print(pidx.value_counts())
#  359410                 1
#  359409                 1
# -9223372036854775808    1
# dtype: int64

[jreback]

this doesn't seem necessary and is very confusing, just use dropna indicator

[sinhrks]

Is it OK to always drop iNaT even if input is normal int64?

[jreback]

u know the dtype of the index so u can so this only if ts a date like (period or DatetimeIndex)
eg look in the very next elif

[jreback]

and u can simply mask the results (if dropna) then box

[sinhrks]

OK, how about this?

[jreback]

better

[jreback]

rebase this

[jreback]

@sinhrks pls rebase

[sinhrks]

@jreback Rebased.

[jreback]

can you run a perf test on this?

[sinhrks]

Attached. replace looks irrelevant for this.

reshape_stack_simple                         |   4.2387 |   3.6786 |   1.1522 |
append_frame_single_mixed                    |   3.7584 |   3.1943 |   1.1766 |
replace_replacena                            |   3.1474 |   2.6023 |   1.2094 |
replace_fillna                               |   4.2147 |   3.1803 |   1.3252 |
-------------------------------------------------------------------------------
Test name                                    | head[ms] | base[ms] |  ratio   |
-------------------------------------------------------------------------------

Ratio < 1.0 means the target commit is faster then the baseline.
Seed used: 1234

Target [01a9e18] : BUG: DTI.value_counts doesnt preserve tz
Base   [bfa9bc5] : Merge pull request #7733 from sinhrks/tsplot_bug

[jreback]

yep, agreed. ping when green

[sinhrks]

@jreback Now green

[jreback]

didn't want to combine with the regular value_counts (in the OpsMixin)? (too hard?)

[sinhrks]

Possible. Actually current logic has if-else clause, not much difference putting it to IndexOpsMixin...

[jreback]

k...just trying to avoid dup code as much as possible...lmk

[sinhrks]

Yep, fixed. Pls check.

[jreback]

perfect, ping when green