Support `keep_key` in `Grouper`?

IterKeyZipper has an option to keep the key that was zipped on:

https://github.com/pytorch/data/blob/2cf1f208e76301f3e013b7569df0d75275f1aaee/torchdata/datapipes/iter/util/combining.py#L53

Is this something we want to support going forward? If yes, it would be nice to have this also on Grouper and possibly other similar datapipes. That would come in handy in situations if the key is used multiple times for example if we have a IterKeyZipper after an Grouper.

Additional Context for New Contributors

See comment below

[NicolasHug]

I took a quick look at this issue. I wonder if it would make sense to instead avoid the overlap between keep_key and merge_fn? For example we could remove the keep_key parameter, and instead let the merge_fn function accept key, source_dp_item, ref_dp_item ? Then passing keep_key=True would be equivalent to passing

merge_fn = lambda key, source_dp_item, ref_dp_item: (key, source_dp_item, ref_dp_item)

[ejguan]

Perhaps I’m missing something but it looks like merge_fn isn’t really necessary since it can be achieved just as easily with a subsequent call to Mapper?

Yeah. When IterKeyZipper starts to support multiple datapipes, the merge_fn is not really useful. IIRC, merge_fn is used as a syntax sugar for a sequence of IterKeyZipper.