Add DatetimeAccessor for accessing datetime fields via .dt attribute (

#1356)

* Add DatetimeAccessor for accessing datetime fields via `.dt` attribute

* Cleaning up unit tests

* Cleaning up comments and warnings in accessors

* Indirectly access pandas tslib through Series accessors

* Re-factor injection of datetime field accessor properties

* Undo loop/injection of _get_date_field accessors

* Remove public-facing dt property

* Remove extra 'field' argument from _tslib_field_accessor

* Added support for dask arrays

* Added dask test cases
Fixed a bug where data wasn't computed in correct order

* Simplified _get_date_field for both dask/numpy arrays; additional code review cleanups

* Fixing flake8 complaints

* Adding whats-new entry

* Updated timeseries docs with note about dt accessor

* Moved season accessor to DatetimeAccessor

* Re-factor virtual variable logic to lean on DateTimeAccessor

* Added "Returns" documentation to _get_date_field
Fixed imports to facilitate more direct implementation of DateTimeAccessor as a property in DataArray
Moved _access_through_series to a top-level function in accessors.py so that dask serialization will hopefully work a bit better

* Adding timestamp accessor

* Hard-coding expected dtypes for each datetime field

* Fix typo in non-datetime virtual variable access

* Update What's New and timeseries docs

doc/time-series.rst

@@ -88,7 +88,22 @@ For more details, read the pandas documentation.
 Datetime components
 -------------------
 
-xarray supports a notion of "virtual" or "derived" coordinates for
+Similar `to pandas`_, the components of datetime objects contained in a
+given ``DataArray`` can be quickly computed using a special ``.dt`` accessor.
+
+.. _to pandas: http://pandas.pydata.org/pandas-docs/stable/basics.html#basics-dt-accessors
+
+.. ipython:: python
+
+    time = time = pd.date_range('2000-01-01', freq='6H', periods=365 * 4)
+    ds = xr.Dataset({'foo': ('time', np.arange(365 * 24)), 'time': time})
+    ds.time.dt.hour
+    ds.time.dt.dayofweek
+
+The ``.dt`` accessor works on both coordinate dimensions as well as
+multi-dimensional data.
+
+xarray also supports a notion of "virtual" or "derived" coordinates for
 `datetime components`__ implemented by pandas, including "year", "month",
 "day", "hour", "minute", "second", "dayofyear", "week", "dayofweek", "weekday"
 and "quarter":
@@ -100,11 +115,13 @@ __ http://pandas.pydata.org/pandas-docs/stable/api.html#time-date-components
     ds['time.month']
     ds['time.dayofyear']
 
-xarray adds ``'season'`` to the list of datetime components supported by pandas:
+For use as a derived coordinate, xarray adds ``'season'`` to the list of
+datetime components supported by pandas:
 
 .. ipython:: python
 
     ds['time.season']
+    ds['time'].dt.season
 
 The set of valid seasons consists of 'DJF', 'MAM', 'JJA' and 'SON', labeled by
 the first letters of the corresponding months.
@@ -124,7 +141,7 @@ calculate the mean by time of day:
 
 For upsampling or downsampling temporal resolutions, xarray offers a
 :py:meth:`~xarray.Dataset.resample` method building on the core functionality
-offered by the pandas method of the same name. Resample uses essentialy the
+offered by the pandas method of the same name. Resample uses essentially the
 same api as ``resample`` `in pandas`_.
 
 .. _in pandas: http://pandas.pydata.org/pandas-docs/stable/timeseries.html#up-and-downsampling

doc/whats-new.rst

@@ -18,6 +18,10 @@ What's New
 v0.9.6 (unreleased)
 -------------------
 
+- Add ``.dt`` accessor to DataArrays for computing datetime-like properties
+for the values they contain, similar to ``pandas.Series`` (:issue:`358`).
+By `Daniel Rothenberg <https://github.com/darothen>`_.
+
 Enhancements
 ~~~~~~~~~~~~
 

xarray/__init__.py

@@ -1,6 +1,7 @@
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
+
 from .core.alignment import align, broadcast, broadcast_arrays
 from .core.common import full_like, zeros_like, ones_like
 from .core.combine import concat, auto_combine

xarray/core/accessors.py

@@ -0,0 +1,150 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from .common import is_datetime_like
+from .pycompat import dask_array_type
+
+from functools import partial
+
+import numpy as np
+import pandas as pd
+
+
+def _season_from_months(months):
+    """Compute season (DJF, MAM, JJA, SON) from month ordinal
+    """
+    # TODO: Move "season" accessor upstream into pandas
+    seasons = np.array(['DJF', 'MAM', 'JJA', 'SON'])
+    months = np.asarray(months)
+    return seasons[(months // 3) % 4]
+
+
+def _access_through_series(values, name):
+    """Coerce an array of datetime-like values to a pandas Series and
+    access requested datetime component
+    """
+    values_as_series = pd.Series(values.ravel())
+    if name == "season":
+        months = values_as_series.dt.month.values
+        field_values = _season_from_months(months)
+    else:
+        field_values = getattr(values_as_series.dt, name).values
+    return field_values.reshape(values.shape)
+
+
+def _get_date_field(values, name, dtype):
+    """Indirectly access pandas' libts.get_date_field by wrapping data
+    as a Series and calling through `.dt` attribute.
+
+    Parameters
+    ----------
+    values : np.ndarray or dask.array-like
+        Array-like container of datetime-like values
+    name : str
+        Name of datetime field to access
+    dtype : dtype-like
+        dtype for output date field values
+
+    Returns
+    -------
+    datetime_fields : same type as values
+        Array-like of datetime fields accessed for each element in values
+
+    """
+    if isinstance(values, dask_array_type):
+        from dask.array import map_blocks
+        return map_blocks(_access_through_series,
+                          values, name, dtype=dtype)
+    else:
+        return _access_through_series(values, name)
+
+
+class DatetimeAccessor(object):
+    """Access datetime fields for DataArrays with datetime-like dtypes.
+
+     Similar to pandas, fields can be accessed through the `.dt` attribute
+     for applicable DataArrays:
+
+        >>> ds = xarray.Dataset({'time': pd.date_range(start='2000/01/01',
+        ...                                            freq='D', periods=100)})
+        >>> ds.time.dt
+        <xarray.core.accessors.DatetimeAccessor at 0x10c369f60>
+        >>> ds.time.dt.dayofyear[:5]
+        <xarray.DataArray 'dayofyear' (time: 5)>
+        array([1, 2, 3, 4, 5], dtype=int32)
+        Coordinates:
+          * time     (time) datetime64[ns] 2000-01-01 2000-01-02 2000-01-03 ...
+
+     All of the pandas fields are accessible here. Note that these fields are
+     not calendar-aware; if your datetimes are encoded with a non-Gregorian
+     calendar (e.g. a 360-day calendar) using netcdftime, then some fields like
+     `dayofyear` may not be accurate.
+
+     """
+    def __init__(self, xarray_obj):
+        if not is_datetime_like(xarray_obj.dtype):
+            raise TypeError("'dt' accessor only available for "
+                            "DataArray with datetime64 or timedelta64 dtype")
+        self._obj = xarray_obj
+
+    def _tslib_field_accessor(name, docstring=None, dtype=None):
+        def f(self, dtype=dtype):
+            if dtype is None:
+                dtype = self._obj.dtype
+            obj_type = type(self._obj)
+            result = _get_date_field(self._obj.data, name, dtype)
+            return obj_type(result, name=name,
+                            coords=self._obj.coords, dims=self._obj.dims)
+
+        f.__name__ = name
+        f.__doc__ = docstring
+        return property(f)
+
+    year = _tslib_field_accessor('year', "The year of the datetime", np.int64)
+    month = _tslib_field_accessor(
+        'month', "The month as January=1, December=12", np.int64
+    )
+    day = _tslib_field_accessor('day', "The days of the datetime", np.int64)
+    hour = _tslib_field_accessor('hour', "The hours of the datetime", np.int64)
+    minute = _tslib_field_accessor(
+        'minute', "The minutes of the datetime", np.int64
+    )
+    second = _tslib_field_accessor(
+        'second', "The seconds of the datetime", np.int64
+    )
+    microsecond = _tslib_field_accessor(
+        'microsecond', "The microseconds of the datetime", np.int64
+    )
+    nanosecond = _tslib_field_accessor(
+        'nanosecond', "The nanoseconds of the datetime", np.int64
+    )
+    weekofyear = _tslib_field_accessor(
+        'weekofyear', "The week ordinal of the year", np.int64
+    )
+    week = weekofyear
+    dayofweek = _tslib_field_accessor(
+        'dayofweek', "The day of the week with Monday=0, Sunday=6", np.int64
+    )
+    weekday = dayofweek
+
+    weekday_name = _tslib_field_accessor(
+        'weekday_name', "The name of day in a week (ex: Friday)", object
+    )
+
+    dayofyear = _tslib_field_accessor(
+        'dayofyear', "The ordinal day of the year", np.int64
+    )
+    quarter = _tslib_field_accessor('quarter', "The quarter of the date")
+    days_in_month = _tslib_field_accessor(
+        'days_in_month', "The number of days in the month", np.int64
+    )
+    daysinmonth = days_in_month
+
+    season = _tslib_field_accessor(
+        "season", "Season of the year (ex: DJF)", object
+    )
+
+    time = _tslib_field_accessor(
+        "time", "Timestamps corresponding to datetimes", object
+    )

xarray/core/common.py

@@ -761,3 +761,10 @@ def ones_like(other, dtype=None):
     """Shorthand for full_like(other, 1, dtype)
     """
     return full_like(other, 1, dtype)
+
+
+def is_datetime_like(dtype):
+    """Check if a dtype is a subclass of the numpy datetime types
+    """
+    return (np.issubdtype(dtype, np.datetime64) or
+            np.issubdtype(dtype, np.timedelta64))

xarray/core/dataarray.py

@@ -15,6 +15,7 @@
 from . import rolling
 from . import ops
 from . import utils
+from .accessors import DatetimeAccessor
 from .alignment import align, reindex_like_indexers
 from .common import AbstractArray, BaseDataObject
 from .coordinates import (DataArrayCoordinates, LevelCoordinatesSource,
@@ -158,6 +159,7 @@ class DataArray(AbstractArray, BaseDataObject):
     """
     _groupby_cls = groupby.DataArrayGroupBy
     _rolling_cls = rolling.DataArrayRolling
+    dt = property(DatetimeAccessor)
 
     def __init__(self, data, coords=None, dims=None, name=None,
                  attrs=None, encoding=None, fastpath=False):

xarray/core/dataset.py

@@ -21,7 +21,7 @@
 from .. import conventions
 from .alignment import align
 from .coordinates import DatasetCoordinates, LevelCoordinatesSource, Indexes
-from .common import ImplementsDatasetReduce, BaseDataObject
+from .common import ImplementsDatasetReduce, BaseDataObject, is_datetime_like
 from .merge import (dataset_update_method, dataset_merge_method,
                     merge_data_and_coords)
 from .utils import (Frozen, SortedKeysDict, maybe_wrap_array, hashable,
@@ -32,6 +32,8 @@
                        integer_types, dask_array_type, range)
 from .options import OPTIONS
 
+import xarray as xr
+
 # list of attributes of pd.DatetimeIndex that are ndarrays of time info
 _DATETIMEINDEX_COMPONENTS = ['year', 'month', 'day', 'hour', 'minute',
                              'second', 'microsecond', 'nanosecond', 'date',
@@ -74,20 +76,11 @@ def _get_virtual_variable(variables, key, level_vars=None, dim_sizes=None):
         virtual_var = ref_var
         var_name = key
     else:
-        if ref_var.ndim == 1:
-            date = ref_var.to_index()
-        elif ref_var.ndim == 0:
-            date = pd.Timestamp(ref_var.values)
-        else:
-            raise KeyError(key)
-
-        if var_name == 'season':
-            # TODO: move 'season' into pandas itself
-            seasons = np.array(['DJF', 'MAM', 'JJA', 'SON'])
-            month = date.month
-            data = seasons[(month // 3) % 4]
+        if is_datetime_like(ref_var.dtype):
+            ref_var = xr.DataArray(ref_var)
+            data = getattr(ref_var.dt, var_name).data
         else:
-            data = getattr(date, var_name)
+            data = getattr(ref_var, var_name).data
         virtual_var = Variable(ref_var.dims, data)
 
     return ref_name, var_name, virtual_var

xarray/tests/test_accessors.py

@@ -0,0 +1,96 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import xarray as xr
+import numpy as np
+import pandas as pd
+
+from . import TestCase, requires_dask
+
+
+class TestDatetimeAccessor(TestCase):
+    def setUp(self):
+        nt = 100
+        data = np.random.rand(10, 10, nt)
+        lons = np.linspace(0, 11, 10)
+        lats = np.linspace(0, 20, 10)
+        self.times = pd.date_range(start="2000/01/01", freq='H', periods=nt)
+
+        self.data = xr.DataArray(data, coords=[lons, lats, self.times],
+                                 dims=['lon', 'lat', 'time'], name='data')
+
+        self.times_arr = np.random.choice(self.times, size=(10, 10, nt))
+        self.times_data = xr.DataArray(self.times_arr,
+                                       coords=[lons, lats, self.times],
+                                       dims=['lon', 'lat', 'time'],
+                                       name='data')
+
+    def test_field_access(self):
+        years = xr.DataArray(self.times.year, name='year',
+                             coords=[self.times, ], dims=['time', ])
+        months = xr.DataArray(self.times.month, name='month',
+                              coords=[self.times, ], dims=['time', ])
+        days = xr.DataArray(self.times.day, name='day',
+                            coords=[self.times, ], dims=['time', ])
+        hours = xr.DataArray(self.times.hour, name='hour',
+                             coords=[self.times, ], dims=['time', ])
+
+        self.assertDataArrayEqual(years, self.data.time.dt.year)
+        self.assertDataArrayEqual(months, self.data.time.dt.month)
+        self.assertDataArrayEqual(days, self.data.time.dt.day)
+        self.assertDataArrayEqual(hours, self.data.time.dt.hour)
+
+    def test_not_datetime_type(self):
+        nontime_data = self.data.copy()
+        int_data = np.arange(len(self.data.time)).astype('int8')
+        nontime_data['time'].values = int_data
+        with self.assertRaisesRegexp(TypeError, 'dt'):
+            nontime_data.time.dt
+
+    @requires_dask
+    def test_dask_field_access(self):
+        import dask.array as da
+
+        years = self.times_data.dt.year
+        months = self.times_data.dt.month
+        hours = self.times_data.dt.hour
+        days = self.times_data.dt.day
+
+        dask_times_arr = da.from_array(self.times_arr, chunks=(5, 5, 50))
+        dask_times_2d = xr.DataArray(dask_times_arr,
+                                     coords=self.data.coords,
+                                     dims=self.data.dims,
+                                     name='data')
+        dask_year = dask_times_2d.dt.year
+        dask_month = dask_times_2d.dt.month
+        dask_day = dask_times_2d.dt.day
+        dask_hour = dask_times_2d.dt.hour
+
+        # Test that the data isn't eagerly evaluated
+        assert isinstance(dask_year.data, da.Array)
+        assert isinstance(dask_month.data, da.Array)
+        assert isinstance(dask_day.data, da.Array)
+        assert isinstance(dask_hour.data, da.Array)
+
+        # Double check that outcome chunksize is unchanged
+        dask_chunks = dask_times_2d.chunks
+        self.assertEqual(dask_year.data.chunks, dask_chunks)
+        self.assertEqual(dask_month.data.chunks, dask_chunks)
+        self.assertEqual(dask_day.data.chunks, dask_chunks)
+        self.assertEqual(dask_hour.data.chunks, dask_chunks)
+
+        # Check the actual output from the accessors
+        self.assertDataArrayEqual(years, dask_year.compute())
+        self.assertDataArrayEqual(months, dask_month.compute())
+        self.assertDataArrayEqual(days, dask_day.compute())
+        self.assertDataArrayEqual(hours, dask_hour.compute())
+
+    def test_seasons(self):
+        dates = pd.date_range(start="2000/01/01", freq="M", periods=12)
+        dates = xr.DataArray(dates)
+        seasons = ["DJF", "DJF", "MAM", "MAM", "MAM", "JJA", "JJA", "JJA",
+                   "SON", "SON", "SON", "DJF"]
+        seasons = xr.DataArray(seasons)
+
+        self.assertArrayEqual(seasons.values, dates.dt.season.values)