Scoped defaults

[Divelix]

I want to reuse augmentations/transforms configs for multiple datasets in torch ConcatDataset. For single dataset it would look the following:

defaults:
  - transforms: my_aug
_target_: src.datasets.MyDataset

But for ConcatDataset I would need list of datasets:

defaults:
  - transforms@tf1: my_aug1
  - transforms@tf2: my_aug2
_target_: torch.utils.data.ConcatDataset
datasets:
  - _target_: src.datasets.MyDataset
    transforms: ${...tf1}
  - _target_: src.datasets.MyDataset
    transforms: ${...tf2}

This implementation for ConcatDataset forks fine in itself, but it breaks when used in recursive instantiation, because it returns not just "target", but ["tf1", "tf2", "target"] instead.

Intuitive solution to that problem would be scoped defaults that syntactically could be defined by underscore so my config looked like:

defaults:
  - transforms@_tf1: my_aug1
  - transforms@_tf2: my_aug2
_target_: torch.utils.data.ConcatDataset
datasets:
  - _target_: src.datasets.MyDataset
    transforms: ${_tf1}
  - _target_: src.datasets.MyDataset
    transforms: ${_tf2}

May be I missed smth in docs and this behaviour is already achievable with current hydra functionality, point me to the right place then, please.