Make constructing slices lazily. #1994
Conversation
xarray/core/rolling.py
Outdated
|
|
||
| self._setup_windows() | ||
| self.window_labels = self.obj[self.dim] | ||
| self._stops = np.arange(1, len(self.window_labels) + 1) |
There was a problem hiding this comment.
can we do this only when iteration is requested? that would be a bit more efficient for the common case where iteration is not done.
My guess is that iteration is slow enough that the overhead of recreating these objects will not be noticeable.
|
The benchmark does not need to be complete, but it would be good to ensure we don't have a regression here. |
| se = self.da_long.to_series() | ||
| getattr(se.rolling(window=window), func)() | ||
| else: | ||
| getattr(self.da_long.rolling(x=window), func)() |
There was a problem hiding this comment.
Also added a benchmark using pandas for long 1d-array, in order to make it easier to find the cause of the regression.
But is it too verbose as pandas might make the similar benchmark tests?
| se = self.da_long.to_series() | ||
| getattr(se.rolling(window=window), func)() | ||
| else: | ||
| getattr(self.da_long.rolling(x=window), func)() |
xarray/core/rolling.py
Outdated
|
|
||
| def __iter__(self): | ||
| for (label, indices) in zip(self.window_labels, self.window_indices): | ||
| window = self.obj.isel(**{self.dim: indices}) | ||
| _stops = np.arange(1, len(self.window_labels) + 1) |
There was a problem hiding this comment.
nit: no need to preface with an underscore for variables that are already limited in scope to this method.
whats-new.rstfor all changes.Quick fix of #1993.
With this fix, the script shown in #1993 runs
Bottleneck: 0.08317923545837402 s
Pandas: 1.3338768482208252 s
Xarray: 1.1349339485168457 s