DataArray .rolling() unclear behaviour when center=False

[juliencarponcy]

What is your issue?

Hi,

I am using the rolling().construct() method which I found very convenient and efficient.

I had timeseries with 2 dimensions: time and channel. I used the construct to produce small overlapping windows of samples on all channels:

xr_data['full_windowed_eeg'] = xr_data['resampled_eeg'] \
							.rolling(resampled_time=window_size, min_periods=None) \
							.construct("window_tvec", stride=1, keep_attrs=True) \
							.dropna('resampled_time') \
							.rename({'resampled_time':'window_time'}).copy()

However, after not obtaining the result I expected, I found out that the new coord window_time, was corresponding to the original time coords at the end/right of the window, and not as the first time coord of the window as I expected.

There is no argument to specify this apparently, as in its current state, it allows only for taking the "center coord" or the "right coord" (if center=False).

I expect that the way I wanted it is not so uncommon, so implementing that possibility would be great. But more urgently, I would find it extremely useful and avoiding puzzling debbugging if this behaviour was clearly explained in the API reference.

But maybe I am missing something here?

Thanks for your great work !

[welcome]

Thanks for opening your first issue here at xarray! Be sure to follow the issue template!
If you have an idea for a solution, we would really welcome a Pull Request with proposed changes.
See the Contributing Guide for more.
It may take us a while to respond here, but we really value your contribution. Contributors like you help make xarray better.
Thank you!

[dcherian]

I think you're being surprised by the fact that rolling.construct pads with window-1 NaNs at the start of the dimension. This behaviour is not configurable at the moment.

[juliencarponcy]

Indeed, I saw it and get rid of it at one point with dropna(). I guess it is worth mentioning / expliciting a little bit in the docs.
The puzzling behaviour, is that for plotting, as per my previous code I have to indroduce an .isel(<3rd_dim>=-1) instead of just .isel(<3rd_dim>=0) if I want to plot signal aligned. More annoying is that if I want to fetch in another DataArray (here 'full_code') based on the newly created time_vector, it fetches data offseted by the window_size. To correct, I'll probably have to introduce a .shift({'window_time': -window_size}) on the time vector created by .construct()

For the plots, I remedy with .isel(<3rd_dim>=-1) instead of just .isel(<3rd_dim>=0) as stated before:

from matplotlib import pyplot as plt
t_before_s = 0.5
t_after_s = 0.5
time_slice = slice(xr_test_data.epoch_onsets[0] + pd.Timedelta(-t_before_s,'s'), xr_test_data.epoch_onsets[0] + pd.Timedelta(t_after_s,'s'))

plt.figure()
xr_test_data['raw_eeg'].sel(time=time_slice, channel='O1').plot.line(x='time', hue='channel')
plt.figure()
xr_test_data['resampled_eeg'].sel(resampled_time=time_slice, channel='O1').plot.line(x='resampled_time', hue='channel')
plt.figure()
xr_test_data['full_code'].sel(resampled_time=time_slice).plot.line(x='resampled_time')
plt.figure()
xr_test_data['windowed_eeg'].sel(window_time=time_slice, channel='O1').isel(window_tvec=-1).plot.line()

I do not feel confident enough to explicit the behaviour of construct() and specify examples by myself for the docs, but will gladly assist and contribute if someone feels like it.