GitLab

* improve perf on convert_image_dtype and add tests

* add reference tests

* use bitshifts for int to int

* revert bitshifts for int to int upscale

* fix warning ignore

* pin pyav to <10

* pin av in GHA workflows as well

* also pin in M1 workflow

* fix, ignore, or assert warnings for consistency tests

* fix, ignore, or assert warnings for kernel infos

* fix to_image_tensor for numpy inputs

* make image from numpy contiguous

* fix test

* Add raft_stereo weights

* Update the metrics layout

* [Nova][WIP] Add Linux CPU Unittests for Torchvision

* use conda-builder image since conda installation is needed

* install torch dep with conda instead

* use circleCI command to run tests

* larger instance to avoid OOM issues

* proper syntax for self-hosted runners

* 4xlarge instance

* 8xlarge

* 12xlarge

* use setup-miniconda job

* add back PATH change to help setup py detect conda

* run conda shell script

* install other deps up front

* git config and undo path change

* revert to local conda install

* conda-builder image

* support for whole python version matrix

* clean up the conda env once we are done with the job

* Use softmax in _get_gaussian_kernel1d

* Revert "Use softmax in _get_gaussian_kernel1d"

This reverts commit eb8fba36302d2da9e06e6f40afaaf901b276a771.

* Code update

* Relaxed tolerance in consistency tests for GaussianBlur and ElasticTransform

* Code review updates

* Update test_prototype_transforms_consistency.py

* [proto] Reduce number of calls of __torch_function__

* Use DisableTorchFunction and super

* Use self._tensor

* Fixes mypy and color space handling

* revert Image.new_like

* WIP

* Perf opt with ref to tensor and properties

* Removed requires_grad property

* Use _tensor ref

* Revert "Use _tensor ref"

This reverts commit 38f8e21242830fed46ddf31287edb67c1abd124a.

* Update torchvision/prototype/features/_feature.py
Co-authored-by: Philip Meier <github.pmeier@posteo.de>
Co-authored-by: Philip Meier <github.pmeier@posteo.de>

* Fix `ToDtype` to avoid errors when a type is not defined.

* Nit `(features.is_simple_tensor, features._Feature)` to `(Tensor,)`

* Fixing linter

* Adding comment.

* Switch back to indexing. Python's default dict seems to have a nasty behaviour.

* Extend RandomShortestSize to support Video specific flavour of the augmentation

* Adding a test.

* Apply changes from code review

* introduce _check method for type checks on prototype transforms

* cleanup

* Update torchvision/prototype/transforms/_geometry.py
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* introduce _check on new transforms

* _check -> _check_inputs

* always check inputs in _RandomApplyTransform
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* introduce nearest-exact interpolation

* update prototype tests

* update stable tests

* expand ToDtype to support multiple conversions at once

* simplify

Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* [proto] Performance improvements for equalize op

* Added tests

* Removing unnecessary methods/classes.

* Unions instead of ImageOrVideo types

* Fixing JIT issue.

* Revert "Add seeds on Kernel Info and reduce randomness for Gaussian Blur (#6741)"

This reverts commit 6e72f2fd

.

* add fixture to fix the RNG seed

* re-add changes to gaussian_blur_* sample input shapes
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* [proto] Small improvement for tensor equalize op

* Fix code formatting

* Added a comment on the ops

* Increase inactivity timeout for binary build jobs

* Fix binary build steo

* Add seeds on Kernel Info and reduce randomness for Gaussian Blur

* Fix linter

* Change `image_size` to `spatial_size`

* Fix linter

* Fixing more tests.

* Adding get_num_channels_video and get_spatial_size_* kernels for video, masks and bboxes.

* Refactor get_spatial_size

* Reduce the usage of `query_chw` where possible

* Rename `query_chw` to `query_spatial_size`

* Adding `get_num_frames` dispatcher and kernel.

* Adding jit-scriptability tests

* Use the variable name sizes instead of scales for consistency

* Add the missing type hints

* Restore the naming back to scales instead of sizes to avoid backwards incompatibility

* Add video support on MixUp and CutMix

* Switch back to roll

* Fix tests and mypy

* Another mypy fix

* Adding support of Video to missed Transforms and Kernels

* Fixing Grayscale Transform.

* Fixing FiveCrop and TenCrop Transforms.

* Fix Linter

* Fix more kernels.

* Add `five_crop_video` and `ten_crop_video` kernels

* Added a TODO.

* Missed Video isinstance

* nits

* Fix bug on AugMix

* Nits and TODOs.

* Reapply Philip's recommendation

* Fix mypy and JIT

* Fixing test

* Remove performance workaround for mask resize

* Fix linter

* bug fixes

* remove unnecessary import

* Fixing linter

* make args and kwargs in ArgsKwargs more accessible

* refactor KernelInfo and DispatcherInfo

* remove ArgsKwargs __getitem__ shortcut again

* enable arbitrary batch size for all prototype kernels

* put back perspective dispatcher

Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* Rewrite `get_dimensions`, `get_num_channels` and `get_spatial_size`

* Remove `get_chw`

* Remove comments

* Make `get_spatial_size` support non-image input

* Reduce the unnecessary use of `get_dimensions*`

* Fix linters

* Fix merge bug

* Linter

* Fix linter

* replace new_like with wrap_like

* fix videos

* revert casting in favor of ignoring mypy