GitLab

Changes image size convention to (h, w). I don't think this is used by
models other than SSD which assumes this convention.
Co-authored-by: Nicolas Granger <nicolas.granger@cea.fr>
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* S3D weight deployment

* Update accuracies.

* Address review comments.

* S3D initial commit

* add model builder code and docstrings

* change classifier submodule, populate weights enum

* fix change of block args from List[List[int]] to ints

* add VideoClassification to transforms

* edit weights url for testing, add s3d to models.video init

* norm_layer changes

* norm_layer and args fix

* Overwrite default dropout

* Remove docs from internal submodules.

* Fix tests

* Adding documentation.

* Link doc from main models.rst

* Fix min_temporal_size

* Adding crop/resize parameters in references script

* Release weights.

* Refactor dropout.

* Adding the weights table in the doc
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>
Co-authored-by: Vasilis Vryniotis <vvryniotis@fb.com>

* init submit

* fix typo

* support ufmt and mypy

* fix 2 unittest errors

* fix ufmt issue

* Apply suggestions from code review
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* unify codes

* fix meshgrid indexing

* fix a bug

* fix type check

* add type_annotation

* add slow model

* fix device issue

* fix ufmt issue

* add expect pickle file

* fix jit script issue

* fix type check

* keep consistent argument order

* add support for pretrained_window_size

* avoid code duplication

* a better code reuse

* update window_size argument

* make permute and flatten operations modular

* add PatchMergingV2

* modify expect.pkl

* use None as default argument value

* fix type check

* fix indent

* fix window_size (temporarily)

* remove "v2_" related prefix and add v2 builder

* remove v2 builder

* keep default value consistent with official repo

* deprecate dropout

* deprecate pretrained_window_size

* fix dynamic padding edge case

* remove unused imports

* remove doc modification

* Revert "deprecate dropout"

This reverts commit 8a13f932815ae25655c07430d52929f86b1ca479.

* Revert "fix dynamic padding edge case"

This reverts commit 1c7579cb1bd7bf2f0f94907f39bee6ed707a97a8.

* remove unused kwargs

* add downsample docs

* revert block default value

* revert argument order change

* explicitly specify start_dim

* add small and base variants

* add expect files and slow_models

* Add model weights and documentation for swin v2

* fix lint

* fix end of files line
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>
Co-authored-by: Joao Gomes <jdsgomes@fb.com>

* Extending to support MViTv2

* Fix docs, mypy and linter

* Refactor the relative positional code.

* Code refactoring.

* Rename vars.

* Update docs.

* Replace assert with exception.

* Updat docs.

* Minor refactoring.

* Remove the square input limitation.

* Moving methods around.

* Modify the shortcut in the attention layer.

* Add ported weights.

* Introduce a `residual_cls` config on the attention layer.

* Make the patch_embed kernel/padding/stride configurable.

* Apply changes from code-review.

* Remove stale todo.

* Expose on Hub the public methods of the registration API

* Limit methods and update docs.

* cleaning up box decoding

* minor nits

* cleanup for box encoding also addded.

* Model registration mechanism.

* Add overwrite options to the dataset prototype registration mechanism.

* Adding example models.

* Fix module filtering

* Fix linter

* Fix docs

* Make name optional if same as model builder

* Apply updates from code-review.

* fix minor bug

* Adding getter for model weight enum

* Support both strings and callables on get_model_weight.

* linter fixes

* Fixing mypy.

* Renaming `get_model_weight` to `get_model_weights`

* Registering all classification models.

* Registering all video models.

* Registering all detection models.

* Registering all optical flow models.

* Fixing mypy.

* Registering all segmentation models.

* Registering all quantization models.

* Fixing linter

* Registering all prototype depth perception models.

* Adding tests and updating existing tests.

* Fix linters

* Fix tests.

* Add beta annotation on docs.

* Fix tests.

* Apply changes from code-review.

* Adding documentation.

* Fix docs.

* intial structure

* fixed types of few variables

* remove the commented code

* list -> List

* encode method will take input as tensors instead of list of tensor
Co-authored-by: Joao Gomes <jdsgomes@fb.com>

* upgrade usort to

* Also update black

* Actually use 1.0.2

* Apply pre-commit
Co-authored-by: Nicolas Hug <contact@nicolas-hug.com>

* fix inplace mutation

* Different attn shouldn't share the same attribute

* a simpler solution
Co-authored-by: YosuaMichael <yosuamichaelm@gmail.com>

* Add ensembled video accuracy on video reference script

* Change the parser func to be similar with classification reference

* Fix typo type->dtype

* Use custom kinetics

* Fix dataset to not getting start_pts

* Change dataset name, and put video_idx at the back

* Ufmt format

* Use functional softmax, updating meta and use it to overwrite eval param

* Fix typo

* Put the eval parameters on the docs for now

* Change meta for video resnet to use frame-rate 15, also change wording on docs

* basic structure

* added constrains

* fixed errors

* thanks to vadim!

* addressing the comments and added docstrign

* Apply suggestions from code review
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* Adding MViT v2 architecture (#6105)

* Adding mvitv2 architecture

* Fixing memory issues on tests and minor refactorings.

* Adding input validation

* Adding docs and minor refactoring

* Add `min_temporal_size` in the supported meta-data.

* Switch Tuple[int, int, int] with List[int] to support easier the 2D case

* Adding more docs and references

* Change naming conventions of classes to follow the same pattern as MobileNetV3

* Fix test breakage.

* Update todos

* Performance optimizations.

* Add support to MViT v1 (#6179)

* Switch implementation to v1 variant.

* Fix docs

* Adding back a v2 pseudovariant

* Changing the way the network are configured.

* Temporarily removing v2

* Adding weights.

* Expand _squeeze/_unsqueeze to support arbitrary dims.

* Update references script.

* Fix tests.

* Fixing frames and preprocessing.

* Fix std/mean values in transforms.

* Add permanent Dropout and update the weights.

* Update accuracies.

* Fix documentation

* Remove unnecessary expected file.

* Skip big model test

* Rewrite the configuration logic to reduce LOC.

* Fix mypy

* Add rough raft-stereo implementation on prototype/models

* Add standard raft_stereo builder, and modify context_encoder to be more similar with original implementation

* Follow original implementation on pre-convolve context

* Fix to make sure we can load original implementation weight and got same output

* reusing component from raft

* Make the raft_stereo_fast able to load original weight implementation

* Format with ufmt and update some comment

* Use raft FlowHead

* clean up comments

* Remove unnecessary import and use ufmt format

* Add __all__ and more docs for RaftStereo class

* Only accept param and not module for raft stereo builder

* Cleanup comment

* Adding typing to raft_stereo

* Update some of raft code and reuse on raft stereo

* Use bool instead of int

* Make standard raft_stereo model jit scriptable

* Make the function _make_out_layer using boolean with_block and init the block_layer with identity

* Separate corr_block into two modules for pyramid and building corr features

* Use tuple if input is not variable size, also remove default value if using List

* Format using ufmt and update ConvGRU to not inherit from raft in order to satisfy both jit script and mypy

* Change RaftStereo docs input type

* Ufmt format raft

* revert back convgru to see mypy errors, add test for jit and fx, make the model fx compatible

* ufmt format

* Specify device for new tensor, dont init module then overwrite and put if-else instead

* Ignore mypy problem on override, put back num_iters on forward

* Revert some effort to make it fx compatible but unnecessary now

* refactor code and remove num_iters from RaftStereo constructor

* Change to raft_stereo_realtime, and specify device directly for tensor creation

* Add description for raft_stereo_realtime

* Update the test for raft_stereo

* Fix raft stereo prototype test to properly test jit script

* Ufmt format

* Test against expected file, change name from raft_stereo to raft_stereo_builder to prevent import error

* Revert __init__.py changes

* Add default value for non-list param on model builder

* Add checking on out_with_block length, add more docs on the encoder

* Use base instead of basic since it is more commonly used

* rename expect file to base as well

* rename on test

* Revert the revert of __init__.py, also revert the adding default value to _raft_stereo to follow the standard pattern

* ufmt format __init__.py

* Add new .. betastatus:: directive to document Beta APIs

* Also add it for the fine-grained video API

* Add directive for all builders and pages of Detection module

* Also segmentation and video models

* Add missing `_version` to the MLPBlock

* fix linter

* Use List[int] instead of int for window_size and shift_size

* Make PatchMerging and SwinTransformerBlock able to handle 2d and 3d cases

* Separate patch embedding from SwinTransformer and enable to get model without head by specifying num_heads=None

* Dont use if before padding so it is fx friendly

* Put the handling on window_size edge cases on separate function and wrap with torch.fx.wrap so it is excluded from tracing

* Update the weight url to the converted weight with new structure

* Update the accuracy of swin_transformer

* Change assert to Exception and nit

* Make num_classes optional

* Add typing output for _fix_window_and_shift_size function

* init head to None to make it jit scriptable

* Revert the change to make num_classes optional

* Revert unneccesarry changes that might be risky

* Remove self.head declaration

* To resolve issue #5964

Add note for resnet architecture

* Update resnet.py

* Update resnet.py

* Update resnet.rst

* Fix stylings

* Add the same notes on model builders

* Improve description

* Apply the change everywhere

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

* Adding an MLP block.

* Adding documentation

* Update typos.

* Fix inplace for Dropout.

* Apply recommendations from code review.

* Making changes on pre-trained models.

* Fix linter

* add swin_s and swin_b variants

* fix swin_b params

* fix n parameters and acc numbers

* adding missing acc numbers

* apply ufmt

* Updating `_docs` to reflect training recipe

* Fix exted for swin_b
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* Doc revamp for optical flow models

* Some more

* Classif models

* Detection

* Segmentation

* quantization

* Video

* optical flow

* tests

* Fix docs

* Fix Video dataset

* Consistency for RAFT dataset names

* use ImageNet-1K

* Use COCO-val2017-VOC-labels for segmentation

* formatting

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

* Adding docs for quantized models.

* Adding docs for video models.

* Adding docs for segmentation models.

* Adding docs for optical flow models.

* Adding docs for detection models.

* Fix typo.

* Make changes from code-review.

* Improving the auto-gen doc.

* Adding details for AlexNet, ConvNext, DenseNet, EfficientNets, GoogLeNet and InceptionV3.

* Fixing location of `_docs`

* Adding docs in the remaining classification models.

* Fix linter

* update paper link for FCOS refence

* remove 'updated version'
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* docs: Added quantized ResNeXt to new docs

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