This is a meta-issue to keep track of all the things we need to do before the branch cut / release. Core branch cut is planned for Monday 13 15th.

This superseeds #7092 and includes stuff unrelated to transformsV2. @pmeier @vfdev-5 please feel free to directly edit this issue with updates, and add anything I may have missed

Before branch cut - Friday 17th

Transforms V2 Stuff

• clamp bounding boxes in some geometry kernels #7215 @pmeier
• Add tests for class transforms (type preservation, input / output format preservation (tuple, dict), pass-through non-native types). @pmeier add proper smoke test for prototype transforms #7238
  • Add more tests for bounding boxes #7276
• Fixed issues with dtype in geom functional transforms v2 #7211 @vfdev-5
• Compatibility layer between stable datasets and prototype transforms #6663 @pmeier
  • Decide on name and location for dataset wrapper - Added wrap_dataset_for_transforms_v2 into datasets and handled beta w… #7279
• Address all TODOs in [NOMERGE] drop in transforms v2 into the v1 tests #7159 @pmeier @vfdev-5 @NicolasHug
  • Fix fill parameter compatibilty with torchscript Fix some annotations in transforms v2 for JIT v1 compatibility #7252 @pmeier
  • add error for max_size with size sequence in resize #7253
  • Ten crop annotation #7254
  • allow integer parameters in ColorJitter #7255
  • Make RandomApply torchscriptable in V2 #7256
  • Fixed issues in elastic transform #7257
  • make fill defaultdict an implementation detail #7258
  • Make rotate_image_tensor's results fully BC with v1 #7271
• make sure there's no critical TODOs left in [NOMRG] TransformsV2 questions / comments #7092 and [NOMRG] TransformsV2 TODOs #7082 @NicolasHug
• Avoid removing BBoxes on transforms, just mark them as degenerate Add SanitizeBoundingBoxes transform #7246 @NicolasHug
  • Make RandomIoUCrop compatible with SanitizeBoundingBoxes #7268
• Add tests for transform presets, and various fixes #7223 @NicolasHug
• Confirm training references - Prototype references  #7220 @NicolasHug (Done, just haven't posted results yets)
• make Image(some_pil_image) work (description: Compatibility layer between stable datasets and prototype transforms #6663 (comment)), @vfdev-5 @pmeier, Image and Mask can accept PIL images #7231
• Figure out a better name than convert_dtype https://github.com/pytorch/vision/pull/7092/files#diff-0c3dc07ddb80b9a83bf674efe1dcaee6aaabab59210d4916d2e35536c498710aR441
• Actually migrate to torchvision.transforms.v2 (or whatever else if we change our mind) Rollout planning for transforms v2 #7097 @pmeier @vfdev-5 @NicolasHug Promote prototype transforms to beta status #7261
  • Double check what we expose as public
  • Split tests for transforms v2 and prototype #7278
• remove datapoints compatibility for prototype datasets #7154 @pmeier
• Figure out why CI didn't trigger on wrong kernel call from datapoint: clamp bounding boxes in some geometry kernels #7215 (comment) @pmeier add tests for perspective start- / endpoints #7226

Other stuff

• Change default of antialias parameter from None to 'warn' #7160 @NicolasHug
• Revert removal and deprecation of int values for interpolation? #7205 @pmeier
  • and add int support to V2 Support integer values for interpolation in the prototype transforms #7248 @pmeier
• Fix Dropout in SwinTransformer unable to set training=False #7103 (CC @oke-aditya :) )
• ONNX test failures #7247 @pmeier @NicolasHug

If time allows

• Decide if we want to remove None as option to antialias
• Properly investigate and address "CUDA error: invalid argument" on model test for vit_h_14 #7143
• Now that Add uint8 support for interpolate for CPU images pytorch#90771 is merged, let Resize() rely on interpolate()'s native uint8 handling instead of converting to and from float.
• Add Dice Loss #6960
• Bikeshed on whether format and spatial_size are adequate meta-data names for bboxes https://github.com/pytorch/vision/pull/7092/files#r1071959354
• Decide whether it's worth enforcing that only one BoundingBox instance is present per sample. Some transforms enforce that, some don't. It could simplify some checks - sort of related: https://github.com/pytorch/vision/pull/7092/files#r1071600864
• Make the dataset wrapper work properly on subclasses of builtin datasets. @pmeier allow subclasses in dataset wrappers #7236
• Make datapoints.FillType and the others private @pmeier make type alias private #7266

After branch cut, before release

• List out all the current design decision that may change and / or that we actively need user feedback on @NicolasHug [FEEDBACK] TransformsV2: What may change in the future (we need your input!) #7319
• Write docs for V2 transforms:
  • Docstrings
  • Tutorial(s)
• Write summary of design changes @pmeier @NicolasHug Recent changes to transforms v2 #7384
• Update previously released content (video, blog posts etc.) to reflect latest changes (typically the V2 are now JIT-scriptable, there's no Label anymore, etc.)
• Write release notes @NicolasHug
• Address recent warnings in the test suite #7265 - Fixed uncaught warnings in tests v2 #7330

Just after release

• Port training references to V2? Would it require releasing new models?