I've been a little slow to update this during the holidays, but we just started looking into updating to 2.7.

I know a few things broke when @birdstream tried it in this forum thread, so will likely require some changes: https://forum.perceptilabs.com/t/boost-your-augmentation-with-a-custom-layer/780

As long as we're on the 2.x branch of TF, is there any reason to think that new TF releases shouldn't work? i.e. could PL have a well-defined and fairly rapid validation process that would allow new releases to be supported asap?

and, Arpan Mathew, just for info, what parts of what's new in TF 2.7 (see https://blog.tensorflow.org/2021/11/whats-new-in-tensorflow-27.html) really interest you? I think the new traceback, error messages could be v. helpful generally, but I would like also to think that, outside customisation, PL would shield us from issues that need them