**Goal**: make it stable by eoy. That means having a "very close to stable" status for 0.16 release.

**RC cut**: 31/08 (28/08 for core). That's in less than 2 months.
**Release**: 04/10

Before release
--------------

- Reach out to other projects: detectron2? HuggingFace? TIMM? fastai?

Ideally before release
---------------------

- Finish the refactored tests
- Add fine-grained RNG support - EDIT: lots of efforts in this direction i.e. https://github.com/pytorch/vision/pull/7848 and https://github.com/pytorch/pytorch/pull/107131. Unfortunately PR in core had to be reverted due to SEV + other reports. This won't happen for the 0.16 release.
- Add a "per-sample RNG" transform wrapper to have individual RNG for each samples in a batch?

Done
----

- Add CutMix/MixUp support (we should proabbly stop calling them batch transforms, this leads to confusion) https://github.com/pytorch/vision/pull/7731
- Submit 10min talk to Pytorch Conf in October 2023
- Fix and clean UX for dtype conversion scaling [issue](https://github.com/pytorch/vision/issues/7756), [PR](https://github.com/pytorch/vision/pull/7759)
- Figure out status of support for batches:
- We **do not** support batches where the number of bboxes varies per image.
- That means we can't have batches for detection.
- TensorDict might be a promising container [for batching bboxes](https://github.com/pytorch/vision/issues/7763)
- Rename BoundingBox into BoundingBoxes https://github.com/pytorch/vision/pull/7778
- Remove use of term `spatial_size` https://github.com/pytorch/vision/pull/7734
- Allow "others" key as a catch-all in dictionary parameters https://github.com/pytorch/vision/issues/7765 - https://github.com/pytorch/vision/pull/7779
- Enable smooth support for user-defined datapoints
- and remove Datapoint methods
- Full revamp of dispatching logic in transforms and functionals https://github.com/pytorch/vision/pull/7747 https://github.com/pytorch/vision/pull/7797 https://github.com/pytorch/vision/pull/7804
- tutorial: https://github.com/pytorch/vision/pull/7795
- dispatcher -> functional renaming https://github.com/pytorch/vision/issues/7798 https://github.com/pytorch/vision/pull/7829
- Remove wrap_like class method and add datapoints.wrap() function https://github.com/pytorch/vision/issues/7803 https://github.com/pytorch/vision/pull/7832
- Enable smooth support for user-defined transforms https://github.com/pytorch/vision/pull/7795
- "enforce" one instance per sample https://github.com/pytorch/vision/pull/7786
- Add `set_return_type` CM (Allow users to choose whether to return Datapoint subclasses or pure Tensor) https://github.com/pytorch/vision/pull/7825
- Add Collab link to all examples https://github.com/pytorch/vision/issues/7794
- resuscitate https://github.com/pytorch/vision/pull/7159 for testing - done, remaining failures are benign
- Add SegMask and DetMask as subclasses. Decided note to do it because we don't agree on what the behaviour should be, and did't have time to align before branch cut https://github.com/pytorch/vision/pull/7909#issuecomment-1700831707
- Add tests for utils https://github.com/pytorch/vision/issues/7785 - EDIT: tests no added still, but we made those private.
- add V2 as a backend to the training references https://github.com/pytorch/vision/pull/7629 https://github.com/pytorch/vision/pull/7724 https://github.com/pytorch/vision/pull/7732 https://github.com/pytorch/vision/pull/7743
- Tons of doc improvements https://github.com/pytorch/vision/pull/7906 https://github.com/pytorch/vision/pull/7889 https://github.com/pytorch/vision/pull/7886 https://github.com/pytorch/vision/pull/7883 https://github.com/pytorch/vision/pull/7881 https://github.com/pytorch/vision/pull/7870 https://github.com/pytorch/vision/pull/7859 https://github.com/pytorch/vision/pull/7849 https://github.com/pytorch/vision/pull/7843 https://github.com/pytorch/vision/pull/7830 https://github.com/pytorch/vision/pull/7795 https://github.com/pytorch/vision/pull/7772

#### Decision points on 13/07

- Review/decide on items from https://github.com/pytorch/vision/issues/7319
- Subclass (un)wrapping: not enough feedback to change anything, keep as-is
- Bounding box clamping: not enough feedback to change anything, keep as-is
- Enforce a single BoundingBox instance in all transforms: Probably best to enforce/assume 1 BBox only. Nicolas to submit a PR for further discussion.
- How to handle labels: OK to keep `labels_getter` param in various transforms that need it (Sanitize, CutMix/MixUp).
- We can still consider storing labels as meta-data, but we need to see it in action if we're going to consider that seriously. Various issues were identified with that previously (what does the meta-data look like when labels are "one-hot" after CutMix/Mixup?)
- How to smoothly support "pairwise" transforms: decided to go ahead and release them as "collate" fn. [PR](https://github.com/pytorch/vision/pull/7731)
- Supporting user-defined datapoints and datapoints methods:
- making the methods private avoids namespace collision, and avoids exposing another way of transforming.
- we still want to expose a "public" dev interface for devs to implement their own datapoints
- considering using a dispatch registration mechanism, Philip to submit PR / POC, we can further discuss pros and cons there
- Tensor pass-through heuristic: Undecided for now
- Pros of simplifying: makes code much easier. Maybe faster??
- Cons: different mental model for users than just "pure tensor == Image". Affects users more directly than the more complex heuristic
- Both heuristics are fully BC in the strict sense.
- Suggestion: see if simplified heuristic also simplifies transform authoring experience (Philip to submit POC / design doc)
- Address any outstanding feedback from [Main Feedback issue](https://github.com/pytorch/vision/issues/6753). EDIT: there's none for now (July 13).