Added the DTD dataset

[pmeier]

This adds the DTD dataset. At its core it is a ImageFolder dataset. Next to a split into train, val and test sets, it contains 10 predefined folds of each partial set.

I've sub-classed ImageFolder class and added support to download it from scratch. I can verify the downloaded archive, but I couldn't think of a way to do this for the extracted content. Thus, the download flag defaults to False.

The index of each image within self.imgs depends on split and fold. I've implemented a converter, which converts the indices between the complete and partial sets. Since the images are only loaded at runtime, split and fold can also be changed at runtime, which triggers a re-initialization of the converter.

[codecov-io]

Codecov Report

Merging #743 into master will increase coverage by 0.22%.
The diff coverage is 25.71%.

@@            Coverage Diff             @@
##           master     #743      +/-   ##
==========================================
+ Coverage   39.72%   39.95%   +0.22%     
==========================================
  Files          29       30       +1     
  Lines        2724     2816      +92     
  Branches      430      437       +7     
==========================================
+ Hits         1082     1125      +43     
- Misses       1568     1615      +47     
- Partials       74       76       +2

Impacted Files	Coverage Δ
torchvision/datasets/__init__.py	100% <100%> (ø)	⬆️
torchvision/datasets/dtd.py	24.63% <24.63%> (ø)
torchvision/models/alexnet.py	38.09% <0%> (-4.02%)	⬇️
torchvision/models/vgg.py	18.18% <0%> (-0.38%)	⬇️
torchvision/models/inception.py	14.41% <0%> (-0.07%)	⬇️
torchvision/datasets/mnist.py	29.19% <0%> (+0.81%)	⬆️
torchvision/transforms/transforms.py	83.41% <0%> (+3.93%)	⬆️

---

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 680c745...4bb9f6d. Read the comment docs.

[pmeier]

I've split the implementation into the overall DTD dataset and subsets which are characterized by the split and fold. Furthermore I removed the ability to change the split and fold at runtime, since its probably mess up the iteration if changed within an epoch.

[fmassa]

Hi,

This looks generally good, thanks!

There are a few linter errors, can you fix them?

./torchvision/datasets/dtd.py:138:72: E128 continuation line under-indented for visual indent
./torchvision/datasets/dtd.py:147:1: W391 blank line at end of file

[pmeier]

Closing this as the DTD dataset does not fit our current understanding of which datasets should be added to torchvision.