[RFC] Add some utility functions to Cityscapes

[TheCodez]

This adds some utility functions to the Cityscapes dataset. Those are functions I use all the time when training FCNs on Cityscapes so it might be useful to have them here directly.

[codecov-io]

Codecov Report

Merging #1025 into master will decrease coverage by 0.24%.
The diff coverage is 26.47%.

@@            Coverage Diff             @@
##           master    #1025      +/-   ##
==========================================
- Coverage   63.29%   63.04%   -0.25%     
==========================================
  Files          65       65              
  Lines        5152     5185      +33     
  Branches      772      782      +10     
==========================================
+ Hits         3261     3269       +8     
- Misses       1667     1692      +25     
  Partials      224      224

Impacted Files	Coverage Δ
torchvision/datasets/cityscapes.py	20.95% <26.47%> (+1.5%)	⬆️

---

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 2694a5d...39a7a2f. Read the comment docs.

[fmassa]

Sure, why not!

But having custom methods to individual classes is a bit annoying in the following sense:

• derived datasets (like ConcatDataset or Subset) do not carry those methods
• not all datasets have these methods, so it requires digging into each individual dataset to discover those attributes.

I do think that we should probably add some more methods to the datasets (like class_names etc), but those depend on the type of the dataset itself.

For example, if we define a ClassificationDataset, where it's clear that it always returns an image and a class label, we can add some more structure to it.

I'm not 100% clear on what the scope of those datasets would be (would Cityscapes be a detection dataset? a segmentation dataset? I don't yet know), but I think adding arbitrary methods to each dataset individually now might imply that in the future we will end up deprecating them.

Thoughts?

[TheCodez]

@fmassa Cityscapes would be a segmentation dataset in like 99% of the use cases.

adding arbitrary methods to each dataset individually now might imply that in the future we will end up deprecating them.

That's a really good point. We shouldn't end up adding all sorts of methods to the classes just to end up deprecating them later.

So closing this for now.