Test for SVHN

[pmeier]

This moves the cast from numpy.ndarray to PIL.Image from __getitem__() to __init__() for the CIFAR10, CIFAR10, and SVHN dataset. Additionally, this adds a dataset test for the SVHN dataset.

[codecov-io]

Codecov Report

Merging #1086 into master will increase coverage by 0.21%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #1086      +/-   ##
==========================================
+ Coverage   64.66%   64.88%   +0.21%     
==========================================
  Files          68       68              
  Lines        5406     5411       +5     
  Branches      829      831       +2     
==========================================
+ Hits         3496     3511      +15     
+ Misses       1661     1643      -18     
- Partials      249      257       +8

Impacted Files	Coverage Δ
torchvision/transforms/transforms.py	80.55% <0%> (-1.24%)	⬇️
torchvision/models/densenet.py	86.79% <0%> (ø)	⬆️
torchvision/models/mnasnet.py	82.71% <0%> (ø)	⬆️
torchvision/models/vgg.py	89.55% <0%> (ø)	⬆️
torchvision/models/resnet.py	88.27% <0%> (ø)	⬆️
torchvision/models/detection/transform.py	78.3% <0%> (+0.2%)	⬆️
torchvision/datasets/svhn.py	67.3% <0%> (+32.69%)	⬆️

---

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 7d8cc19...a90324a. Read the comment docs.

[fmassa]

Thanks a lot for the PR!

The new tests look great!

I have a question regarding passing the creation of the PIL images in the constructor: does it speed up significantly the __getitem__ step? Because this is a potentially breaking change, as users could have accessed the .data field beforehand.
So it would be good to know how much slowdown this represents.

[pmeier]

So it would be good to know how much slowdown this represents.

It takes ~80µs per image (on my machine) for the conversion, thus making the performance advantage negligible. I would argue that the performance was never the decisive factor for this PR. I simply found it odd that we keep the data as numpy.ndarray while we internally use PIL as image backend. But then I'm from a prototypy world and keep forgetting that BC could be more important.

If you think BC is more important here, I'm happy to roll back the changes and only keep the test for SVHN.

[fmassa]

@pmeier I think it's important to keep BC until we have a clear story about what are the fields of datasets that users can use, and those that they shouldn't rely on. Your issue on unifying datasets is a great start, and I'll make some comments today on it.

If you could revert the BC changes and keep only the test for SVHN for now, this would be ready to merge

[fmassa]

Thanks a lot!