Add Caltech101, Caltech256, and CelebA

[bryanhe]

No description provided.

[codecov-io]

Codecov Report

Merging #775 into master will decrease coverage by 2.85%.
The diff coverage is 17.92%.

@@            Coverage Diff             @@
##           master     #775      +/-   ##
==========================================
- Coverage   40.99%   38.13%   -2.86%     
==========================================
  Files          29       32       +3     
  Lines        2747     3126     +379     
  Branches      432      487      +55     
==========================================
+ Hits         1126     1192      +66     
- Misses       1542     1855     +313     
  Partials       79       79

Impacted Files	Coverage Δ
torchvision/datasets/__init__.py	100% <100%> (ø)	⬆️
torchvision/datasets/celeba.py	14.73% <14.73%> (ø)
torchvision/datasets/caltech.py	21.13% <21.13%> (ø)
torchvision/datasets/utils.py	35.1% <9.67%> (-13.29%)	⬇️
torchvision/models/__init__.py	100% <0%> (ø)	⬆️
torchvision/models/googlenet.py	15.87% <0%> (ø)

---

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 6cc4970...5bf4ecf. Read the comment docs.

[fmassa]

Hi,

Thanks for the PR and sorry for the delay in reviewing it.

It generally looks good, I have a few comments.

Also, in a more general question: is Caltech101 and Caltech256 still used for papers nowadays? If I remember properly, you could get very good classification results by just classifying on the average color in the image.

[bryanhe]

I think all of the issues should be resolved now.

I don't think many papers only use Caltech101/256, but I've seen it used for domain adaptation experiments in somewhat recent papers, such as http://papers.nips.cc/paper/6757-learning-multiple-tasks-with-deep-relationship-networks (2017), along with a few papers that it cites.

Do you think some of those (Office-Home and ImageCLEF) are worth including?

[fmassa]

I think this looks almost ready to be merged, thanks!

I have only a few minor nits about using a legacy torch.LongTensor construct, otherwise this is good to be merged

[fmassa]

I haven't myself seen / used Office-Home and ImageCLEF be used much, but if you think it would be valuable to have them in torchvision, if you could send a PR adding them it would be great

[bryanhe]

Last comment is resolved now.

I'll think about whether those two are worth including.
I was also going to add in a few other datasets (DeepFashion and ADE2K are fairly popular, so I assume they're worth adding?). Does it make sense to PR each on separately, or just block them together?

[fmassa]

Separate PRs are better I think.

Thanks!