pd.Grouper support?

[naught101]

In pandas, you can pas a pandas.TimeGrouper object to a .groupby() call, and it allows you to group by month, year, day, or other times, without manually creating a new index with those values first. It would be great if you could do this with xray, but at the moment, I get:

/usr/local/lib/python3.4/dist-packages/xray/core/groupby.py in __init__(self, obj, group, squeeze)
     66             if the dimension is squeezed out.
     67         """
---> 68         if group.ndim != 1:
     69             # TODO: remove this limitation?
     70             raise ValueError('`group` must be 1 dimensional')

AttributeError: 'TimeGrouper' object has no attribute 'ndim'

Not sure how this will work though, because pandas.TimeGrouper doesn't appear to work with multi-index dataframes yet anyway, so maybe there needs to be a feature request over there too, or maybe it's better to implement something from scratch...

[shoyer]

I wrote a resample function last week based on TimeGrouper. See the dev docs for more details: http://xray.readthedocs.org/en/latest/whats-new.html

This should go out in the 0.4.1 release, which I'd like to get out later this week (everyone likes faster release cycles if they are backwards compatible).

It would be pretty straightforward to create some sort of API that gives direct access to the resulting GroupBy object. I was considering something like Dataset.groupby_time. I'd love some input here if you have ideas.

[naught101]

Looks good to me. I don't know enough to be able to comment on the API question.

[shoyer]

Well, I guess the first question is -- are there uses for TimeGrouper that you can't easily do with resample?

I suppose the simplest (no new method) would be to allow passing a dict where the key is the time dimension and the value is the grouper. Something like ds.groupby({'time': pd.TimeGrouper(...)}) or ds.groupby(time=pd.TimeGrouper(...)).

[naught101]

Unfortunately I'm not familiar enough with pd.resample and pd.TeimGrouper to know the difference in what they can do. resample looks like it would cover my use-cases, although the docs are pretty limited, and don't actually specify the format of the rule format...

One thing that I would like to be able to do that is not covered by resample, and might be covered by TimeGrouper is to group over month only (not month and year), in order to create a plot of mean seasonal cycle (at monthly resolution), or similarly, a daily cycle at hourly resolution. I haven't figured out if I can do that with TimeGrouper yet though.

[shoyer]

Indeed, I need to complete the resample docs before doing a release :).

For your other use case, you just want to group by 'time.month' (or 'time.hour'), e.g.,
http://xray.readthedocs.org/en/v0.4/examples/weather-data.html#probability-of-freeze-by-calendar-month

[naught101]

Heh, I meant the pandas docs - they don't specify the rule argument format either

time.month and time.hour do exactly what I need. They aren't mentioned in the docs at http://xray.readthedocs.org/en/stable/groupby.html, and I'm not sure how I'd guess that they exist, so perhaps they should be added to that page? It doesn't appear to be something that exists in pandas..

[shoyer]

For pandas resample, see here: http://pandas.pydata.org/pandas-docs/stable/timeseries.html#up-and-downsampling

The doc string could definitely use an update there, too -- see pandas-dev/pandas#5023 (I think I'll try to update this, too)

For time.month, see this part of the docs:
http://xray.readthedocs.org/en/stable/data-structures.html#dataarray-coordinates

I'm going to consolidate all the time/date functionality into a new documentation page for the next release of xray, since this is kind of all over the place now.

Also, I should probably break up that monolithic page on "Data structures", perhaps into "Basics" and "Advanced" pages.

[naught101]

Ah, cool, thanks for that link, I missed that in the docs.

One thing that would be nice (in both pandas and xray) is a time.timeofday. I can't figure out how to do it with time.hour and time.minute - I need half-hourly resolution averaging. time.time does something in xray, but it seems to never complete, and it doesn't work at all in pandas.

[shoyer]

Hmm. ds['time.time'] appears to be broken in xray (it casts to a datetime64 array). I'll make a new issue for that.

However, it should work in pandas -- you can do df['time'].dt.time:

In [13]: t = pd.date_range('2000-01-01', periods=10000, freq='H')

In [14]: t.time
Out[14]:
array([datetime.time(0, 0), datetime.time(1, 0), datetime.time(2, 0), ...,
       datetime.time(13, 0), datetime.time(14, 0), datetime.time(15, 0)], dtype=object)

The simplest way to do timeofday, though, is probably just to calculate ds.coords['timeofday'] = ds['time.hour'] + 60 * ds['time.minute']. Then you can do ds.groupby('timeofday').

[naught101]

Nice.

Ok, I have hit a stumbling block, and this is much more of a support request, so feel free to direct me else where, but since we're on the topic, I want to do something like:

start = 2002
n_years = 4
new_data = []
for year in range(start, start + n_years):
    days = 365 if year%4 else 365
    for d in range(days):
        day_data = mean + annual_cycle.isel(dayofyear=d) + daily_cycle
        day_data.coords['time'] = datetime.datetime(year,1,1) + datetime.timedelta(day=d, hour=day_data.timeofday.hour, minute=day_data.timeofday.minute)
        new_data.append(day_data)
xray.concat(new_data)

where mean, annual_cycle, and daily_cycle are overall mean, annual cycle at daily resolution, and daily cycle at 30 minute resolution (the latter two bias corrected by subtracting the mean). I'm trying to make a synthetic dataset 4 years long that only includes the mean, seasonal, and daily cycles, but no other variability.

The assignment of day_data['time'] fails because the day_data.timeofday.hour (and .minute) don't work. These are datetime.times, is there an efficient way of converting these to datetime.timedeltas, without first manually taking them out of the DataArray?

[naught101]

Dunno if this is related to the ds['time.time'] problem, but I tried creating the daily_cycle using a pandas.Timedelta as the index (timeofday), but it also appeared to just hang indefinitely when doing the data.groupby('timeofday').mean('time') call..

[naught101]

same problem with numpy.timedelta64 too.

[shoyer]

I don't think the timeofday issue is related to using Timedeltas in the index (and it's certainly not related to the ds['time.time'] issue). Also, it should be exactly equivalent to use timedelta64 or Timedelta -- both should convert to timedelta64 when put in an xray object.

Here's an example that seems to be working properly (except for uselessly display timedeltas in nanoseconds):

In [29]: time = pd.date_range('2000-01-01', freq='H', periods=100)

In [30]: daystart = time.to_period(freq='1D').to_datetime()

In [31]: timeofday = time.values - daystart.values

In [32]: ds = xray.Dataset({'data': ('time', range(100))}, {'time': time, 'timeofday': ('time', timeofday)})

In [33]: ds
Out[33]:
<xray.Dataset>
Dimensions:    (time: 100)
Coordinates:
    timeofday  (time) timedelta64[ns] 0 nanoseconds ...
  * time       (time) datetime64[ns] 2000-01-01 2000-01-01T01:00:00 ...
Data variables:
    data       (time) int64 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 ...

In [34]: ds.groupby('timeofday').mean('time')
Out[34]:
<xray.Dataset>
Dimensions:    (timeofday: 24)
Coordinates:
  * timeofday  (timeofday) timedelta64[ns] 0 nanoseconds ...
Data variables:
    data       (timeofday) float64 48.0 49.0 50.0 51.0 40.0 41.0 42.0 43.0 44.0 45.0 46.0 ...

[naught101]

Ok, weird. That example works for me, but even if I take a really short slice of my data set, the same thing won't work:

In [61]:
d = data.sel(time=slice('2002-01-01','2002-01-03'))
d

Out[61]:

<xray.Dataset>
Dimensions:           (time: 143, timeofday: 70128, x: 1, y: 1, z: 1)
Coordinates:
  * x                 (x) >f8 1.0
  * y                 (y) >f8 1.0
  * z                 (z) >f8 1.0
  * time              (time) datetime64[ns] 2002-01-01T00:30:00 ...
  * timeofday         (timeofday) timedelta64[ns] 1800000000000 nanoseconds ...
Data variables:
    SWdown            (time, y, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 14.58 ...
    Rainf_qc          (time, y, x) float64 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ...
    SWdown_qc         (time, y, x) float64 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ...
    Tair              (time, z, y, x) float64 282.9 282.9 282.7 282.6 282.4 281.7 281.0 ...
    Tair_qc           (time, y, x) float64 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ...
    LWdown            (time, y, x) float64 296.7 297.3 297.3 297.3 297.2 295.9 294.5 ...
    PSurf_qc          (time, y, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ...
    latitude          (y, x) float64 -35.66
    Wind              (time, z, y, x) float64 2.2 2.188 1.9 2.2 2.5 2.5 2.5 2.25 2.0 2.35 ...
    LWdown_qc         (time, y, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ...
    Rainf             (time, y, x) float64 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0 ...
    Qair_qc           (time, y, x) float64 1.0 0.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ...
    longitude         (y, x) float64 148.2
    PSurf             (time, y, x) float64 8.783e+04 8.783e+04 8.782e+04 8.781e+04 ...
    reference_height  (y, x) float64 70.0
    elevation         (y, x) float64 1.2e+03
    Qair              (time, z, y, x) float64 0.00448 0.004608 0.004692 0.004781 ...
    Wind_qc           (time, y, x) float64 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 1.0 ...
Attributes:
    Production_time: 2012-09-27 12:44:42
    Production_source: PALS automated netcdf conversion
    Contact: palshelp@gmail.com
    PALS_fluxtower_template_version: 1.0.2
    PALS_dataset_name: TumbaFluxnet
    PALS_dataset_version: 1.4

In [62]:
d.groupby('timeofday').mean('time')

That last command will not complete - it will run for minutes. Not really sure how to debug that behaviour.

Perhaps it's to do with the long/lat/height variables that really should be coordinates (I'm just using the data as it came, but I can clean that, if necessary)