Support empty target_type for CelebA dataset

[dniku]

Previously, passing target_type=[] wouldn't work because of this line:

celeba = CelebA('whatever/dir', target_type=[])
celeba[0]  # IndexError: list index out of range

This is a workaround that returns only X instead of X, target if target_type is empty.

[codecov-io]

Codecov Report

Merging #1351 into master will decrease coverage by 0.66%.
The diff coverage is 0%.

@@            Coverage Diff            @@
##           master   #1351      +/-   ##
=========================================
- Coverage   65.47%   64.8%   -0.67%     
=========================================
  Files          75      76       +1     
  Lines        5827    5955     +128     
  Branches      892     912      +20     
=========================================
+ Hits         3815    3859      +44     
- Misses       1742    1822      +80     
- Partials      270     274       +4

Impacted Files	Coverage Δ
torchvision/datasets/celeba.py	19.23% <0%> (-0.51%)	⬇️
torchvision/models/_utils.py	67.44% <0%> (-21.02%)	⬇️
torchvision/models/densenet.py	77.85% <0%> (-9.06%)	⬇️
torchvision/datasets/hmdb51.py	27.65% <0%> (ø)	⬆️
torchvision/io/__init__.py	100% <0%> (ø)	⬆️
torchvision/datasets/kinetics.py	34.78% <0%> (ø)	⬆️
torchvision/datasets/ucf101.py	25% <0%> (ø)	⬆️
torchvision/io/_video_opt.py	27.02% <0%> (ø)
torchvision/models/segmentation/_utils.py	80% <0%> (+0.83%)	⬆️

---

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 f677ea3...43414fa. Read the comment docs.

[fmassa]

Thanks for the PR!

What do you think about always returning image, target, but target potentially being None if you pass an empty list?

I'm not sure I'd like to see the return values of datasets to change depending on a constructor argument. This has caused some issues already, like #1273

[dniku]

I had some doubts about the interface: in addition to X and X, None it is also possible to return X, tuple(). I'm not sure which option is best. What would you suggest?

[fmassa]

I'd say just return None. If we always returned tuple I'd say that tuple might be better, but in the case of a single output option this is not the case.

[dniku]

Amended code and docstring. Please re-review.

[fmassa]

I have a comment that I think would minimize the changes in the PR, let me know what you think

[fmassa]

I would maybe simplify the changes in this PR like the following, and only change the following places:

if target:
    target = tuple(target) if len(target) > 1 else target[0]
    
    if self.target_transform is not None:
        target = self.target_transform(target)
else:
    target = None

Thoughts?

[dniku]

I had doubts about keeping the diff small vs keeping the logic straightforward. Amended the commit following your advice.

Also, added a check for a meaningless combination of parameters (empty target_type with nonempty target_transform).

[fmassa]

LGTM, thanks!