ENH: Please provide `register_groupBy_accessor`

[talegari]

Thanks for providing the option to create custom accessors via pandas.api.extensions. This is prefereable method to extend pandas dataframes over inheritance/sub-classing. Currently, this supports these: dataframe, series, index.

Please extend this support for groupBy objects.

[TomAugspurger]

Can you provide some motivation for this? Why doesn't passing callables to agg / transform / apply suffice?

[talegari]

Tom,

Thanks for picking this up.

I started with writing dplyr like operations on pandas dataframes with the primary intent of letting the user not worry about the indexes (see is_tidy_frame). For example, mutate will work on both grouped and ungrouped dataframes via the right accessor.

Having the groupby accessor will avoid two things:

1. Checking whether the input is a tidyframe
2. Avoid the pointless if else in the code. Here is the current code:

def mutate(df, func_list, check = True):

    if check:
        assert is_tidy_frame(df)

    if is_grouped_frame(df):
        groupvars = group_vars(df)
        for func in func_list:
            df = df.apply(func)
        df = df.reset_index(drop = True).pipe(group_by, groupvars)
    else:
        for func in func_list:
            df = df.pipe(func)

    return(df)

[pwwang]

This is definitely helpful for developers who want to wrap pandas APIs.

I had some workarounds to implement datar, which reimages pandas APIs and align them to tidyverse's.