GitLab

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

* factor out find_classes

* use find_classes in video datasets

* adapt old tests

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

* remove unused imports after manual review

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

* remove two more instances
Co-authored-by: Ben Weinstein <benweinstein@Bens-MacBook-Pro.local>
Co-authored-by: Vasilis Vryniotis <datumbox@users.noreply.github.com>

* concatenate small tensors into big ones to reduce the use of shared file descriptor (#1694)

Summary:
Pull Request resolved: https://github.com/pytorch/vision/pull/1694



- PT dataloader forks worker process to speed up the fetching of dataset example.  The recommended way of multiprocess context is `forkserver` rather than `fork`.

- Main process and worker processes will share the dataset class instance, which avoid duplicating the dataset and save memory. In this process, `ForkPickler(..).dumps(...)` will be called to serialize the objects, including objects within dataset instance recursively. `VideoClips` instance internally uses O(N) `torch.Tensor` to store per-video information, such as pts, and possible clips, where N is the No. of videos.

- During dumping, each `torch.Tensor` will use one File Descriptor (FD). The OS default max limit of FD is 65K by using `ulimit -n` to query. The number of tensors in `VideoClips` often exceeds the limit.

- To resolve this issue, we use a few big tensors by concatenating small tensors in the `__getstate__()` method, which will be called during pickling. This will only require O(1) tensors.

- When this diff is landed, we can abondon D19173248

In D19173397, in ClassyVision, we change the mp context from `fork` to `forkserver`, and finally can run the PT dataloader without hanging issues.

Reviewed By: fmassa

Differential Revision: D19179991

fbshipit-source-id: c8716775c7c154aa33d93b25d112d2a59ea688a9

* Try to fix Windows

* Try fix Windows v2

* Disable tests on Windows

* Add back necessary part

* Try fix OSX (and maybe Windows)

* Fix

* Try enabling Windows
Co-authored-by: Zhicheng Yan <zyan3@fb.com>

Now the dataset is not working properly because of this line of code `indices = [i for i in range(len(video_list)) if video_list[i][len(self.root) + 1:] in selected_files]`. 
Performing the `len(self.root) + 1` only make sense if there is no training / to root

```
>>> root = 'data/ucf-101/videos'
>>> video_path = 'data/ucf-101/videos/activity/video.avi'
>>> video_path [len(root ):]
'/activity/video.avi'
>>> video_path [len(root ) + 1:]
'activity/video.avi'
```

Appending the root path also to the selected files is a simple solution and make the dataset works with and without a trailing slash.

* Fix hmdb51.py typo

* Fix ucf101.py typo

* add metadata to video dataset classes. bug fix. more robustness

* query video backend within VideoClips class

* Fix tests

* Fix lint

Fixed video labelling after `subset` call for HMDB51 dataset (hmdb51.py) (EDIT: UCF101 as well) (#1240)

* Fixed video labelling after subset for HMDB51 dataset

* Fixed video labelling after subset for HMDB51 dataset
Co-authored-by: Eric Tang <etang21@stanford.edu>
Co-authored-by: Ryan Cao <ryancao@stanford.edu>

* UCF 101 Labeling fixes

- Analogous fix to HMDB51 to maintain correct labels after the train-test split
- Additional change to the `select_fold` method in `ucf101.py` to correctly reflect the annotation format
Co-authored-by: Ryan Cao <ryancao@stanford.edu>
Co-authored-by: Eric Tang <etang21@stanford.edu>

* Add docs to UCF101 and HMDB51 datasets

* Remove extra line

* Lint

* Add HMDB51 and UCF101

* Remove debug code