Add vggface2 dataset

[jgbradley1]

Closes issue #1139.

There was a PR for this last year but it never got approved and finalized. Since the design of the Dataset classes has changed quite a bit since then, I thought I would submit this more updated version for review.

Notable changes:

• The dataset class uses bbox and landmark annotations that are available on the VGGFace2 website. From what I can tell, these are the only annotations that are available for all identities in the dataset. Other annotation information is available but only for a subset of face identities.
• No preprocessing of the dataset is performed. This option is left up to the user who can freely define transform, target_type and target_transform in order to make whatever changes they want.

cc @fmassa @dakshjotwani

[codecov]

Codecov Report

Merging #2910 (a392514) into master (78159d6) will decrease coverage by 0.39%.
The diff coverage is 21.42%.

@@            Coverage Diff             @@
##           master    #2910      +/-   ##
==========================================
- Coverage   73.39%   73.00%   -0.40%     
==========================================
  Files          99      100       +1     
  Lines        8825     8871      +46     
  Branches     1391     1402      +11     
==========================================
- Hits         6477     6476       -1     
- Misses       1929     1970      +41     
- Partials      419      425       +6     

Impacted Files	Coverage Δ
torchvision/datasets/vggface2.py	20.28% <20.28%> (ø)
torchvision/datasets/__init__.py	100.00% <100.00%> (ø)
torchvision/ops/feature_pyramid_network.py	91.20% <0.00%> (-3.30%)	⬇️
torchvision/models/detection/anchor_utils.py	92.10% <0.00%> (-2.70%)	⬇️
torchvision/ops/deform_conv.py	70.96% <0.00%> (-1.34%)	⬇️
torchvision/models/detection/backbone_utils.py	94.28% <0.00%> (-1.27%)	⬇️
torchvision/models/densenet.py	82.81% <0.00%> (-0.77%)	⬇️
torchvision/models/shufflenetv2.py	86.04% <0.00%> (-0.77%)	⬇️
torchvision/models/detection/_utils.py	85.03% <0.00%> (-0.40%)	⬇️
torchvision/models/inception.py	84.64% <0.00%> (-0.40%)	⬇️
... and 9 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 78159d6...a363104. Read the comment docs.

[dakshjotwani]

Thanks for bringing this PR back to life :)

Can we add test cases for VGGFace2 similar to the imagenet dataset? I remember I was about to do them for my PR before we forgot about it.

[facebook-github-bot]

Hi @jgbradley1!

Thank you for your pull request and welcome to our community. We require contributors to sign our Contributor License Agreement, and we don't seem to have you on file.

In order for us to review and merge your code, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!

[jgbradley1]

Unit tests have been added and tested.

[jgbradley1]

This PR is as ready-to-go as it gets I think. Unless someone has an issue with return types or documentation suggestions, I don't have anything else to add.

[jgbradley1]

Before more work is done on this PR, we should wait to see if the dataset is made available again. Looks like the dataset has been taken down for almost a month.

VGGFace2 citation count: 912

[facebook-github-bot]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Facebook open source project. Thanks!

[facebook-github-bot]

Thank you for signing our Contributor License Agreement. We can now accept your code for this (and any) Facebook open source project. Thanks!

[yassineAlouini]

Thanks for the contribution @jgbradley1 and sorry for the late reply.

This is a new dataset request and will be added to the new dataset API (once it is stable enough).

If you want to/have the time, you can add this contribution to the new dataset or you can leave it and someone else will take the work from here. Notice that once this port is done, this PR will be closed.

Again, thanks for the code!

[pmeier]

To complement @yassineAlouini's comment:

I'm not sure how I missed this PR since it seems complete. I'm very sorry about that @jgbradley1. Unfortunately, we are moving to a new datasets API in the near future and thus, we will no longer accept PRs that add new datasets to the old API. There are two options now:

1. Given that you already put in the work in this PR, we would happily accept a new PR adding this to the new API. You can refer to this guide on how to do that.
2. Someone else could take this over and build on what you have done. In this case I will make sure to credit you properly.

How would you like to go about it @jgbradley1?