date_range treats '.5s' as '5s'

[keszybz]

I'm new to pandas, so maybe I'm missing something, but I'd expected freq='.5s' to give me 500ms intervals, not 5s intervals:

pd.date_range(0, periods=5, freq='.5s')[1]
Timestamp('1970-01-01 00:00:05', offset='5S')
pd.date_range(0, periods=5, freq='500ms')[1]
Timestamp('1970-01-01 00:00:00.500000', offset='500L')
pd.version
'0.14.1'
import sys; sys.version
'3.4.1 (default, Jul 6 2014, 20:01:46) \n[GCC 4.9.0]'

Behaviour is the same under Python 2 and 3.

[jreback]

these don't take fractional units ATM (I don't think was ever implemented). I don't think very difficult to fix. welcome to have a pull-request.

[jbradish]

https://github.com/pydata/pandas/blob/master/pandas/tseries/frequencies.py#L97

I think this may have to do with the integer division when assigned values to a freq group. For example, if you also try .05s, it gets cut to 5S. I believe this is because of the (freq // 1000) operation used in grouping.

Someone else might be able to correct me or give additional insights.

Edit: jreback beat me to it. I might look into this tonight as it seems like it would be interesting, but simple. No promises though.

[jreback]

@jbradish sounds great!

basically you need to patch pandas.tseries.frequencies.to_offset() to handle.

[sinhrks]

I think fix should be 2 parts:

1. offsets.Tick support fraction

Currently, Tick and DateOffset coerces input to int implicitly. I think Tick can return smaller tick to based on fraction, and DateOffset should explicitly raise.

pd.tseries.offsets.Hour(0.5)
# <0 * Hours>

pd.tseries.offsets.MonthEnd(0.5)
# <0 * MonthEnds>

2. to_offset support fraction

After Tick support fraction, it can be done with a slight change of regex and impl (related to #13930).