Dear contributors and users,

I wanted to share an update on what we’ll be prioritizing in TorchVision from now on. In short:

In details:

• The transforms.v2 module, and pre-processing in general, remains our principal area of investments. We’ll keep investing in performance improvements and in new features.
• The utils and ops modules are maintained, but new features will only be added on a best-effort basis.
• The datasets and models modules are still usable and they are not deprecated, but we’ll decrease support. This may mean for example that if a dataset URL breaks, we won’t necessarily actively try to fix the download functionality. The models are already very stable, so regressions are unlikely. New features will likely not be added.
• The capabilities of the io module will eventually move to TorchCodec, where we’re consolidating all of PyTorch encoding/decoding capabilities:
  • For video, changes are effective from now (0.22): the video decoding and encoding capabilities are deprecated and will be removed in 0.24 (end of 2025)!. We strongly encourage you to migrate towards TorchCodec ASAP.
  • The image decoders/encoders remain in torchvision for the time being. They will eventually be migrated to TorchCodec, but there is no ETA on this for now. The migration will involve our usual 2-version deprecation cycle, so you’ll be given a heads-up.

Most of these changes are contributor-facing, not user-facing. As a result, contributors may observe reduced engagement on issues and PRs that relate to the de-prioritized areas (models, datasets, ops and utils).

EDIT: Deleted some comments below which were feature requests and unrelated questions. Feel free to comment on this issue with questions that are in scope, but please open separate issues for feature requests.