Subclassing geodataframe

What is the best way to subclass a GeoDataFrame? I tried to follow some instructions as indicated in this link for subclassing pandas data structures but this does not seem to propagate to the from_file classmethod:

import geopandas

class Hectometer(geopandas.GeoSeries):
    #each instance represents 1 hectometer
    @property
    def _constructor(self):
        return Hectometer

    @property
    def _constructor_expanddim(self):
        return Hectometers

class Hectometers(geopandas.GeoDataFrame):
    @property
    def _constructor(self):
        return Hectometers

    @property
    def _constructor_sliced(self):
        return Hectometer
def test():
    hms = Hectometers.from_file(r'vector\Hectometerraai_A2.shp')
    print(type(hms))
if __name__ == '__main__':
    test()

The above code outputs <class 'geopandas.geodataframe.GeoDataFrame'>, whereas I would like to get the subclassed object instead. I tried about overriding the from_file constructor but this does not seem to refer to self in the original GeoDataFrame class which means it will always return a GeoDataFrame and not its subclass.

What is an elegant way to subclass GeoDataFrames and GeoSeries such that I retrieve the subclassed object from the from_file method?

[m-richards]

@NickLucche I don’t have the same background as Joris to comment within the context of the pandas subclassing. I would also just reiterate the point about the use case to subclass - and whether what you’re doing with subclassing a geodataframe could be achived through composition of a Dataset class containing a geodataframe as an attribute (but if you’ve been using it for months, it’s quite likely you already have a good reason).

From my perspective, I see changing the from_* classmethods as an innocuous change - it would at least help your use case. But there would still be a considerable amount of work to have properly implemented subclassing - because every time _constructor is called internally by pandas, the custom subclass would be lost (which as an aside if you’ve been developing your class wrapper with geopandas 0.10.2, _constructor has changed on the main branch and would be more complicated to feed a class through.

Also there would be a considerable testing burden to cover all cases and noting replacing GeoDataFrame / GeoSeries with cls in every spot will not always do the desired thing (for instance GeoDataFrame.set_geometry is monkey patched onto pd.DataFrame, so if we replace the GeoDataFrame constructor call within, then it will no longer produce GeoDataFrames, but rather dataframes).

[NickLucche]

Hey @m-richards, thanks a lot for your response and your insights! My implementation is indeed based on 0.10.0 and it relies on a custom _constructorto create instances, as suggested (at the time) by the pandas documentation, and I would favor maintaining this behavior of having to provide a custom _constructor to make a subclass. But I see your point in the efforts required to having such changes tested on a feature you didn’t intend to support, so I won’t move forward until testing out a few implementations options based on your feedback.

Nonetheless, I would personally find it useful to have the tiny changes proposed in PR #2433 merged in the main branch, let me know if you have further feedback on that.