pypi / torchvision
torchvision
What changed in torchvision 0.28 to 0.29, from its official release notes, upgrade guides and advisories. Sources checked 2026-10-18.
1 security1 deprecations1 new APIs and options3 behavior and defaults4 fixes and other
See which of these apply to your code, ranked, in your agent:
npx @unstale/cli inittorchvision 0.296 changes
Deprecations 1
The image decoders and encoders in torchvision.io are deprecated and will be removed in a future release; they are available in TorchCodec.
Behavior and default changes 2
The PyTorch requirement is relaxed from an exact 2.14.0 pin to torch>=2.14 because TorchVision is ABI stable with respect to PyTorch 2.14.
TorchVision 0.29 is ABI stable with torch 2.14 and compatible with future torch versions, so upgrading torch will not require a new TorchVision version.
Fixes and other changes 3
Fixed deformable convolution kernels always running on the default stream.
Fixed the JPEG transform for non-contiguous batches.
Fixed gradient overaccumulation in ROI ops.
torchvision 0.284 changes
Security 1
Fixes a malformed-GIF decoder bug that could write outside the allocated tensor's memory.
New APIs and options 1
Resize transforms now accept strings for the interpolation parameter.
Behavior and default changes 1
wrap() preserves metadata for custom TVTensor subclasses.
Fixes and other changes 1
F.resize now honors NEAREST_EXACT for tv_tensors.Mask inputs instead of silently using NEAREST.