GitLab

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

* add tests for HMDB51 dataset

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

* add tests for Kinetics400

* use create_random_string()
Co-authored-by: Philip Meier <pmeier.github@posteo.de>
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* include intended structure of root directory in docstring of Kinetics400

* fix syntax
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: vfdev <vfdev.5@gmail.com>
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: vfdev <vfdev.5@gmail.com>
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Current implementation is generating bad graph after onnx conversion. So replacing with flatten like in mobilenetv3 code.
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* deprecate _video modules

* flake8
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* always use default config as base

* fix test_all_configs decorator

* lint

* add a utility function to create a random string

* move output check of inject_fake_data to dedicated method

* always disable download and extract functionality

Co-authored-by: vfdev <vfdev.5@gmail.com>

* make UsageError an Exception rather than RuntimeError

* separate fake data injection and dataset args handling

* adapt tests for Coco

* fix Coco implementation

* add documentation

* fix VideoDatasetTestCase

* adapt UCF101 tests

* cleanup

* allow FileNotFoundError for test without fake data

* Revert "fix Coco implementation"

This reverts commit e2b693881654b3e2462a73e6d22bb01c1b738f8a.

* lint

* fix UCF101 tests

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* enable default frames per clips for video test cases

* add tests for UCF101

* remove old tests as well as fake data generation

* better explain frames_per_clip overriding

* lint
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Summary:
Pull Request resolved: https://github.com/pytorch/vision/pull/3422

In CUDA-10.1 or below it is a no-op, but few of the tests will start to fail if update to 10.2 or newer, for example:
```
torchvision/transforms/transforms.py", line 1043, in forward
    transformed_tensor = torch.mm(flat_tensor, self.transformation_matrix)
RuntimeError: Deterministic behavior was enabled with either `torch.use_deterministic_algorithms(True)` or `at::Context::setDeterministicAlgorithms(true)`, but this operation is not deterministic because it uses CuBLAS and you have CUDA >= 10.2. To enable deterministic behavior in this case, you must set an environment variable before running your PyTorch application: CUBLAS_WORKSPACE_CONFIG=:4096:8 or CUBLAS_WORKSPACE_CONFIG=:16:8. For more information, go to https://docs.nvidia.com/cuda/cublas/index.html#cublasApi_reproducibility


```

Reviewed By: meyering

Differential Revision: D26563045

fbshipit-source-id: 65e5fabc8cdb471a72f931df55389ed7c0063e54
Co-authored-by: Nikita Shulga <nshulga@fb.com>

* put back

* flake8

* Specify coordinate constraints

* some more

* flake8

* removed call to optimze_for_mobile

* removed unrelated changes

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* use common download utils in VOC and SBDataset

* add tests for VOC

* use common base class for VOC datasets

* remove old voc test and fake data generation
Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

This reverts commit 74de51d6.

* add base class for datasets tests

* add better type hints

* add documentation to subclasses

* add utility functions to create files / folders of random images and videos

* fix imports

* remove class properties

* fix smoke test

* fix type hints

* fix random size generation

* add Caltech256 as example

* add utility function to create grid of combinations

* add CIFAR100? as example

* lint

* add missing import

* improve documentation

* create 1 frame videos by default

* remove obsolete check

* return path of files created with utility functions

* [test] close PIL file handles before deletion

* fix video folder creation

* generalize file handle closing

* fix lazy imports

* add test for transforms

* fix explanation comment

* lint

* force load opened PIL images

* lint

* copy default config to avoid inplace modification

* enable additional arg forwarding

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

Co-authored-by: Francisco Massa <fvsmassa@gmail.com>

* restore cache

* Trigger CI again to see if cache is hit

- Added conda-forge channel

* Avoid freezing bn1 if all layers are trainable.

* Remove misleading comments.

Fixes #3393

* Replace type T with accumulator.

* Upcast tensors of box ops to avoid overflow in multiplications.

Summary:
In opt mode, the linker could strip out the symbols related to the torchvision ops if they had not been explicitly requested.

Adding link-whole is one way of fixing this.

Reviewed By: datumbox, ppwwyyxx

Differential Revision: D26249713

fbshipit-source-id: 4e55e2ebb483922fe308fd8b83456439bde728c9