Easier inheritance of TimeSeries object

I was looking to extend TimeSeries object to make it easier to convert it to other not yet implemented and internal formats. Usually, I’d do something like

from darts import TimeSeries as DartsTimeSeries

class TimeSeries(DartsTimeSeries):
    def to_my_custom_format(self):
        pass

The problem with this approach is that all factory methods in TimeSeries are static and are not aware of the underlying class, which means that

from my_module import TimeSeries
TimeSeries.from_dataframe(df, "time", "value")   # -> darts.timeseries.TimeSeries

still returns darts.timeseries.TimeSeries object without to_my_custom_format() method. Is there a reason why factory methods are implemented as @staticmethod and not @classmethod? The latter would be more friendly from inheritance point of view.

[hrzn]

I think there is still another case for making inheritance easier that is (AFAIK) not yet solved by #536.

In some methods of the TimeSeries class, instantiation is done by explicitly calling TimeSeries. Some examples of methods where this is done: head, tail, but also abs etc. I illustrate how this impedes easier inheritance below:

import pandas as pd
from darts.timeseries import TimeSeries # from feature/classmethods branch

class Child(TimeSeries):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        
df = pd.DataFrame({'time':pd.date_range('20210101', periods=10), 'value': range(10)})
ts = Child.from_dataframe(df, time_col='time')
print(type(ts), type(ts.head()))

The above code will result in: <class ‘main.Child’> <class ‘main.TimeSeries’>

If one would change the current TimeSeries call in the head method to self.__class__ the result then becomes: <class ‘main.Child’> <class ‘main.Child’>

Thus retaining any methods defined in Child.

Good catch. We’ll also address these cases, thanks for reporting.

[hrzn]

Yes, this is a good point. There’s no good reason and we’ll make them classmethods.