GitLab

* Change enum name for weights contributed by community.

* Adding multiweight support to squeezenet.

* Move model builder at the bottom of the file, so we can use proper typing.

* Adding GoogLeNet with multi-weight support.

* Simplify expression.

* Add logging to torchvision ops

* Hack to make torchscript work

* Bugfix

* Bugfix

* Lint

* mypy... let's silence it

* Fighting with mymy

* One more try

Summary:
This diff enables ImageNet to be usable from within torchvision.

It implements a different codepath for some families of datasets that can be sharded into multiple pieces, and uses it if within fbcode.

For now all previously supported datasets are also supported (MNIST, CIFAR and FashionMNIST), in addition to ImageNet.

Pull Request resolved: https://github.com/pytorch/vision/pull/4776

Reviewed By: datumbox

Differential Revision: D31929684

fbshipit-source-id: 58f9fe2ed6ce731a6b43a38be912983247eff562
(cherry picked from commit b43fead8ea5164e37637ade0e80477e94f993365)

* Adding multiweight support for shufflenetv2 prototype models

* Revert "Adding multiweight support for shufflenetv2 prototype models"

This reverts commit 31fadbee7d1a65cd73ae43dfd4ac6e97e7ca7b01.

* Adding multiweight support for shufflenetv2 prototype models

* Revert "Adding multiweight support for shufflenetv2 prototype models"

This reverts commit 4e3d900f796c1e3e667312087e77956ca4a4c017.

* Adding multiweight support for shufflenetv2 prototype models

* addressing PR comment

* Add mutli-weight support for VideoResNet.

* Fix linter.

* Minor refactoring.

* Update comments.

Co-authored-by: Nicolas Hug <nicolashug@fb.com>

* adding multiweight support for deeplabv3 prototype models

* adding default values for optional params

* fixing bug

* addressing PR comment

* Change test_random_apply

* Change test_random_choice

* Change test_randomness

* took care of RandomVert/HorizFlip

* take care of RandomGrayScale

* minor cleanup

* avoid 0 degree rotation just in case

* Fix flakiness on the TestStochasticDepth test.

* Fix minor bug when p=1.0

* Remove device and dtype setting.

* cache dataset info

* try loading categories file by default
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* Adding multi-weight support to LRASPP

* Adding tests for segmentation models.

* Skip segmentation test by default.

Co-authored-by: Prabhat Roy <prabhatroy@fb.com>

* replace np.frombuffer with torch.frombuffer in MNIST prototype

* cleanup

* appease mypy

* more cleanup

* clarify inplace offset

* fix num bytes for floating point data

Summary:
Pull Request resolved: https://github.com/pytorch/vision/pull/4620

Pull Request resolved: https://github.com/pytorch/nestedtensor/pull/455



Fixes missing kernel guards as identified by D30072495

Reviewed By: jingsh, xush6528

Differential Revision: D31553158

fbshipit-source-id: 80de017ba2ddc52e2a684df9b3eae5de84ed49f4
Co-authored-by: Richard Barnes <rbarnes@fb.com>

Summary:
Pull Request resolved: https://github.com/pytorch/vision/pull/4571



Fixing windows build for `torchvision` when `MOBILE` is defined. In `csrc/vision.cpp`, since `PyMODINIT_FUNC` depends on `Python.h` I added the same condition for `PyMODINIT_FUNC` as the one for `import <PyTorch.h>`.

Reviewed By: malfet

Differential Revision: D31488734

fbshipit-source-id: fd52a9ab161288d9c8bf3b9e0ab9da856a4ebadb
Co-authored-by: Mengwei Liu <larryliu@fb.com>

* Update model checkpoint for resnet50.

* Add get_weight method to retrieve weights from name.

* Update the references to support prototype weights.

* Fixing mypy typing.

* Switching to a python 3.6 supported equivalent.

* Add unit-test.

* Add optional num_classes.

Summary: As titled.

Reviewed By: sallysyw

Differential Revision: D31813675

fbshipit-source-id: 8ff879be172530bcc41aae5d71c67e1e90f09cbf
Co-authored-by: Kai Zhang <kaizh@fb.com>

* Expand usage logging to detection models

Summary: Track all detection model instantiation.

Reviewed By: sallysyw

Differential Revision: D31741603

fbshipit-source-id: 89467db5d89ea41c5a9ab95d1846264d0a1f4199

* Fix lint
Co-authored-by: Kai Zhang <kaizh@fb.com>

Summary: We would like to track the internal usage of TorchVision components. We're tracking [datasets usage](https://fburl.com/daiquery/dqpmemn3

) now. This diff expand the tracking to all models.

Reviewed By: fmassa

Differential Revision: D31441632

fbshipit-source-id: e26072e582ac9f832c2056307ebf0eccf2ed6c9c
Co-authored-by: Kai Zhang <kaizh@fb.com>

* Add mutli-weight support for FCN.

* Rename base_size to resize_size.

* Adding multiweight support for mnasnet prototype

* Adressing PR comments

* adding return type to _mnasnet
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* Adding multiweight support for mobilenetv3 prototype

* Update EMA every X iters.

* Adding AdamW optimizer.

* Adjusting EMA decay scheme.

* Support custom weight decay for Normalization layers.

* Fix identation bug.

* Change EMA adjustment.

* Quality of life changes to faciliate testing

* ufmt format

* Fixing imports.

* Adding FixRes improvement.

* Support EMA in store_model_weights.

* Adding interpolation values.

* Change train_crop_size.

* Add interpolation option.

* Removing hardcoded interpolation and sizes from the scripts.

* Fixing linter.

* Incorporating feedback from code review.

* adding efficientnet to prototype

* Removing _efficientnet_conf from prototype and main efficientnet classification models

* fixing merge conflicts

* fixing merge conflicts

* fixing lint errors

* fixing lint errors

* fixing lint errors

* adding efficientnet to prototype

* renaming _efficientnet_model in torchvision.model to be aligned with prototype implementation

* adding specific cop sizes to ech model

* adding specific cop and resize transforms

* fixing lint errors

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

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* rename pretrained model versions
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

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

* adding interpolation model information to all classification prototype models

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

* Migrate mnist dataset from np.frombuffer

* Add a copy with bytearray for non-writable buffers

* Add byte reversal for mnist
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* make mypy more strict for prototype datasets

* fix code format

* apply strictness only to datasets

* fix more mypy issues

* cleanup

* fix mnist annotations

* refactor celeba

* warn on redundant casts

* remove redundant cast

* simplify annotation

* fix import