Rolling window with as_strided

asv_bench/benchmarks/__init__.py

@@ -9,6 +9,14 @@
 _counter = itertools.count()
 
 
+def parameterized(names, params):
+    def decorator(func):
+        func.param_names = names
+        func.params = params
+        return func
+    return decorator
+
+
 def requires_dask():
     try:
         import dask

asv_bench/benchmarks/rolling.py

@@ -0,0 +1,50 @@
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import pandas as pd
+import xarray as xr
+
+from . import parameterized, randn, requires_dask
+
+nx = 3000
+ny = 2000
+nt = 1000
+window = 20
+
+
+class Rolling(object):
+    def setup(self, *args, **kwargs):
+        self.ds = xr.Dataset(
+            {'var1': (('x', 'y'), randn((nx, ny), frac_nan=0.1)),
+             'var2': (('x', 't'), randn((nx, nt))),
+             'var3': (('t', ), randn(nt))},
+            coords={'x': np.arange(nx),
+                    'y': np.linspace(0, 1, ny),
+                    't': pd.date_range('1970-01-01', periods=nt, freq='D'),
+                    'x_coords': ('x', np.linspace(1.1, 2.1, nx))})
+
+    @parameterized(['func', 'center'],
+                   (['mean', 'count'], [True, False]))
+    def time_rolling(self, func, center):
+        getattr(self.ds.rolling(x=window, center=center), func)()
+
+    @parameterized(['window_', 'min_periods'],
+                   ([20, 40], [5, None]))
+    def time_rolling_np(self, window_, min_periods):
+        self.ds.rolling(x=window_, center=False,
+                        min_periods=min_periods).reduce(getattr(np, 'nanmean'))
+
+    @parameterized(['center', 'stride'],
+                   ([True, False], [1, 200]))
+    def time_rolling_construct(self, center, stride):
+        self.ds.rolling(x=window, center=center).construct(
+            'window_dim', stride=stride).mean(dim='window_dim')
+
+
+class RollingDask(Rolling):
+    def setup(self, *args, **kwargs):
+        requires_dask()
+        super(RollingDask, self).setup(**kwargs)
+        self.ds = self.ds.chunk({'x': 100, 'y': 50, 't': 50})

doc/api.rst

@@ -467,6 +467,32 @@ DataArray methods
    DataArray.load
    DataArray.chunk
 
+Rolling objects
+===============
+
+.. autosummary::
+   :toctree: generated/
+
+   core.rolling.DataArrayRolling
+   core.rolling.DataArrayRolling.construct
+   core.rolling.DataArrayRolling.reduce
+   core.rolling.DatasetRolling
+   core.rolling.DatasetRolling.construct
+   core.rolling.DatasetRolling.reduce
+
+GroupByObjects
+==============
+
+.. autosummary::
+   :toctree: generated/
+
+   core.groupby.DataArrayGroupBy
+   core.groupby.DataArrayGroupBy.apply
+   core.groupby.DataArrayGroupBy.reduce
+   core.groupby.DatasetGroupBy
+   core.groupby.DatasetGroupBy.apply
+   core.groupby.DatasetGroupBy.reduce
+
 Plotting
 ========
 

doc/computation.rst

@@ -158,20 +158,42 @@ Aggregation and summary methods can be applied directly to the ``Rolling`` objec
     r.mean()
     r.reduce(np.std)
 
-Note that rolling window aggregations are much faster (both asymptotically and
-because they avoid a loop in Python) when bottleneck_ is installed. Otherwise,
-we fall back to a slower, pure Python implementation.
+Note that rolling window aggregations are faster when bottleneck_ is installed.
 
 .. _bottleneck: https://github.com/kwgoodman/bottleneck/
 
-Finally, we can manually iterate through ``Rolling`` objects:
+We can also manually iterate through ``Rolling`` objects:
 
 .. ipython:: python
 
    @verbatim
    for label, arr_window in r:
       # arr_window is a view of x
 
+Finally, the rolling object has ``construct`` method, which gives a
+view of the original ``DataArray`` with the windowed dimension attached to
+the last position.
+You can use this for more advanced rolling operations, such as strided rolling,
+windowed rolling, convolution, short-time FFT, etc.
+
+.. ipython:: python
+
+    # rolling with 2-point stride
+    rolling_da = r.construct('window_dim', stride=2)
+    rolling_da
+    rolling_da.mean('window_dim', skipna=False)
+
+Because the ``DataArray`` given by ``r.construct('window_dim')`` is a view
+of the original array, it is memory efficient.
+
+.. note::
+  numpy's Nan-aggregation functions such as ``nansum`` copy the original array.
+  In xarray, we internally use these functions in our aggregation methods
+  (such as ``.sum()``) if ``skipna`` argument is not specified or set to True.
+  This means ``rolling_da.mean('window_dim')`` is memory inefficient.
+  To avoid this, use ``skipna=False`` as the above example.
+
+
 .. _compute.broadcasting:
 
 Broadcasting by dimension name

doc/whats-new.rst

@@ -43,6 +43,14 @@ Documentation
 
 Enhancements
 ~~~~~~~~~~~~
+- Improve :py:func:`~xarray.DataArray.rolling` logic.
+  :py:func:`~xarray.DataArrayRolling` object now supports
+  :py:func:`~xarray.DataArrayRolling.construct` method that returns a view
+  of the DataArray / Dataset object with the rolling-window dimension added
+  to the last position. This enables more flexible operation, such as strided
+  rolling, windowed rolling, ND-rolling, and convolution.
+  (:issue:`1831`, :issue:`1142`, :issue:`819`)
+  By `Keisuke Fujii <https://github.com/fujiisoup>`_.
 - Reduce methods such as :py:func:`DataArray.sum()` now handles object-type array.
 
   .. ipython:: python
@@ -106,6 +114,10 @@ Enhancements
 
 Bug fixes
 ~~~~~~~~~
+- Rolling aggregation with ``center=True`` option now gives the same result
+  with pandas including the last element (:issue:`1046`).
+  By `Keisuke Fujii <https://github.com/fujiisoup>`_.
+
 - Support indexing with a 0d-np.ndarray (:issue:`1921`).
   By `Keisuke Fujii <https://github.com/fujiisoup>`_.
 - Added warning in api.py of a netCDF4 bug that occurs when

xarray/core/common.py

@@ -412,6 +412,11 @@ def groupby(self, group, squeeze=True):
         grouped : GroupBy
             A `GroupBy` object patterned after `pandas.GroupBy` that can be
             iterated over in the form of `(unique_value, grouped_array)` pairs.
+
+        See Also
+        --------
+        core.groupby.DataArrayGroupBy
+        core.groupby.DatasetGroupBy
         """
         return self._groupby_cls(self, group, squeeze=squeeze)
 
@@ -471,9 +476,6 @@ def rolling(self, min_periods=None, center=False, **windows):
         """
         Rolling window object.
 
-        Rolling window aggregations are much faster when bottleneck is
-        installed.
-
         Parameters
         ----------
         min_periods : int, default None
@@ -491,7 +493,8 @@ def rolling(self, min_periods=None, center=False, **windows):
 
         Returns
         -------
-        rolling : type of input argument
+        Rolling object (core.rolling.DataArrayRolling for DataArray,
+        core.rolling.DatasetRolling for Dataset.)
 
         Examples
         --------
@@ -519,6 +522,11 @@ def rolling(self, min_periods=None, center=False, **windows):
         array([  1.,   2.,   3.,   4.,   5.,   6.,   7.,   8.,   9.,  10.])
         Coordinates:
           * time     (time) datetime64[ns] 2000-02-15 2000-03-15 2000-04-15 ...
+
+        See Also
+        --------
+        core.rolling.DataArrayRolling
+        core.rolling.DatasetRolling
         """
 
         return self._rolling_cls(self, min_periods=min_periods,

xarray/core/dask_array_ops.py

@@ -1,6 +1,8 @@
-"""Define core operations for xarray objects.
-"""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
 import numpy as np
+from . import nputils
 
 try:
     import dask.array as da
@@ -24,3 +26,70 @@ def dask_rolling_wrapper(moving_func, a, window, min_count=None, axis=-1):
     # trim array
     result = da.ghost.trim_internal(out, depth)
     return result
+
+
+def rolling_window(a, axis, window, center, fill_value):
+    """ Dask's equivalence to np.utils.rolling_window """
+    orig_shape = a.shape
+    # inputs for ghost
+    if axis < 0:
+        axis = a.ndim + axis
+    depth = {d: 0 for d in range(a.ndim)}
+    depth[axis] = int(window / 2)
+
+    offset = 1 if window % 2 == 0 else 0
+
+    # pad the original array before the operation in order to avoid copying
+    # the output array (output array is just a view).
+    if center:
+        start = int(window / 2)  # 10 -> 5,  9 -> 4
+        end = window - 1 - start
+    else:
+        start, end = window - 1, 0
+
+    drop_size = depth[axis] - offset - np.maximum(start, end)
+    if drop_size < 0:
+        # ghosting requires each chunk should be larger than depth.
+        if -drop_size < depth[axis]:
+            pad_size = depth[axis]
+            drop_size = depth[axis] + drop_size
+        else:
+            pad_size = -drop_size
+            drop_size = 0
+        shape = list(a.shape)
+        shape[axis] = pad_size
+        chunks = list(a.chunks)
+        chunks[axis] = (pad_size, )
+        fill_array = da.full(shape, fill_value, dtype=a.dtype, chunks=chunks)
+        a = da.concatenate([fill_array, a], axis=axis)
+
+    if depth[axis] > min(a.chunks[axis]):
+        raise ValueError(
+            "For window size %d, every chunk should be larger than %d, "
+            "but the smallest chunk size is %d. Rechunk your array\n"
+            "with a larger chunk size or a chunk size that\n"
+            "more evenly divides the shape of your array." %
+            (window, depth[axis], min(a.chunks[axis])))
+
+    # We temporary use `reflect` boundary here, but the edge portion is
+    # truncated later.
+    boundary = {d: fill_value for d in range(a.ndim)}
+
+    # create ghosted arrays
+    ag = da.ghost.ghost(a, depth=depth, boundary=boundary)
+
+    # apply rolling func
+    def func(x, window, axis=-1):
+        x = np.asarray(x)
+        rolling = nputils._rolling_window(x, window, axis)
+        return rolling[(slice(None), ) * axis + (slice(offset, None), )]
+
+    chunks = list(a.chunks)
+    chunks.append(window)
+    out = ag.map_blocks(func, dtype=a.dtype, new_axis=a.ndim, chunks=chunks,
+                        window=window, axis=axis)
+
+    # crop the edge points
+    index = (slice(None),) * axis + (slice(drop_size,
+                                           drop_size + orig_shape[axis]), )
+    return out[index]

xarray/core/duck_array_ops.py

@@ -16,7 +16,9 @@
 import pandas as pd
 
 from . import npcompat
+from . import nputils
 from . import dtypes
+from . import dask_array_ops
 from .pycompat import dask_array_type
 from .nputils import nanfirst, nanlast
 
@@ -281,6 +283,10 @@ def f(values, axis=None, skipna=None, **kwargs):
         dtype = kwargs.get('dtype', None)
         values = asarray(values)
 
+        # dask requires dtype argument for object dtype
+        if (values.dtype == 'object' and name in ['sum', ]):
+            kwargs['dtype'] = values.dtype if dtype is None else dtype
+
         if coerce_strings and values.dtype.kind in 'SU':
             values = values.astype(object)
 
@@ -372,3 +378,16 @@ def last(values, axis, skipna=None):
         _fail_on_dask_array_input_skipna(values)
         return nanlast(values, axis)
     return take(values, -1, axis=axis)
+
+
+def rolling_window(array, axis, window, center, fill_value):
+    """
+    Make an ndarray with a rolling window of axis-th dimension.
+    The rolling dimension will be placed at the last dimension.
+    """
+    if isinstance(array, dask_array_type):
+        return dask_array_ops.rolling_window(
+            array, axis, window, center, fill_value)
+    else:  # np.ndarray
+        return nputils.rolling_window(
+            array, axis, window, center, fill_value)

xarray/core/missing.py

@@ -13,6 +13,7 @@
 from .computation import apply_ufunc
 from .utils import is_scalar
 from .npcompat import flip
+from . import rolling
 
 
 class BaseInterpolator(object):
@@ -329,4 +330,8 @@ def _get_valid_fill_mask(arr, dim, limit):
     '''helper function to determine values that can be filled when limit is not
     None'''
     kw = {dim: limit + 1}
-    return arr.isnull().rolling(min_periods=1, **kw).sum() <= limit
+    # we explicitly use construct method to avoid copy.
+    new_dim = rolling._get_new_dimname(arr.dims, '_window')
+    return (arr.isnull().rolling(min_periods=1, **kw)
+            .construct(new_dim, fill_value=False)
+            .sum(new_dim, skipna=False)) <= limit

xarray/core/npcompat.py

@@ -2,6 +2,17 @@
 from __future__ import division
 from __future__ import print_function
 import numpy as np
+from distutils.version import LooseVersion
+
+
+if LooseVersion(np.__version__) >= LooseVersion('1.12'):
+    as_strided = np.lib.stride_tricks.as_strided
+else:
+    def as_strided(x, shape=None, strides=None, subok=False, writeable=True):
+        array = np.lib.stride_tricks.as_strided(x, shape, strides, subok)
+        array.setflags(write=writeable)
+        return array
+
 
 try:
     from numpy import nancumsum, nancumprod, flip