TTA with lazy config

[kretes]

There is a nice support for TTA in detectron2 in https://github.com/facebookresearch/detectron2/blob/cbbc1ce26473cb2a5cc8f58e8ada9ae14cb41052/detectron2/modeling/test_time_augmentation.py

Am I right to see that it only works with old config and not new (lazy) one?

GeneralizedRCNNWithTTA(nn.Module) can just be instantiated using old config.
Furthermore - https://github.com/facebookresearch/detectron2/blob/main/tools/lazyconfig_train_net.py does not mention TTA, and it uses an AMP/Simple trainer which as well doesn't have any TTA mentioned.

In the old config there is a working trainer test_with_tta implementation in https://github.com/facebookresearch/detectron2/blob/main/projects/DensePose/densepose/engine/trainer.py.

1. Am I right that this won't work with lazy config?
2. Is there a way to backport lazy config to a new one (we are training a vitdet model, hence using lazy config) like calling some conversion method?
3. If not - is this list enough that need to be ported? (adapting GeneralizedRCNNWithTTA to work on init variables, and hence being open to usage from both configs + writing some trainer script for that.