davidorme
commented
7 months ago I'd been thinking that the _include structure handles the padding, but of course the _set_times method and other places wants self.datetimes to map onto the first axis. It does seem easier to pad the datetimes than to fix all the downstream special cases.
I've done something similar before using np.unique, which might be useful?
datetimes = np.arange(
np.datetime64('2002-01-03 03:15'),
np.datetime64('2002-01-06 09:45'),
np.timedelta64(30, 'm')
)
# Work out observations by date
dates = datetimes.astype("datetime64[D]")
unique_dates, date_counts = np.unique(dates, return_counts=True)
complete_count = date_counts.max()
padding = (complete_count - date_counts[0], complete_count - date_counts[-1])
if padding != (0,0):
# Do we have at least _one_ complete day
if len(unique_dates) < 3:
ValueError("Not enough data")
# Get a complete set of observation times from day 2 - we have guaranteed that the
# second day is complete.
complete_times = datetimes[dates == unique_dates[1]]
observation_timedeltas = complete_times - complete_times[0].astype("datetime64[D]")
# Recreate the complete timeseries, using an offset to the last observations to avoid an
# off by one error from np.arange.
datetimes = np.arange(
unique_dates[0] + observation_timedeltas[0],
unique_dates[-1] + (observation_timedeltas[-1] + np.timedelta64(1, 's')),
np.diff(observation_timedeltas)[0]
)
MarionBWeinzierl
codecov-commenter
This PR fixes #72 .
It handles potential incomplete first and last days of measurements by padding, both for the datetime array and the value array.
Tasks for this: