Refactor prototype datasets

[pmeier]

This merges a feature branch into main. All changes up to #5777/ cd36d06 were reviewed on individual PRs.

---

The Dataset class now inherits from IterDataPipe rather than being a custom class from which a datapipe is created. This has several upsides:

• The object can have a regular constructor with standard input checking. Before the inputs were checked automatically, but we needed hacky workarounds if not all combinations of parameters were valid. Compare

  vision/torchvision/prototype/datasets/_builtin/voc.py

  Lines 33 to 43 in 5f74f03

  	class VOCDatasetInfo(DatasetInfo):
  	def __init__(self, *args: Any, **kwargs: Any):
  	super().__init__(*args, **kwargs)
  	self._configs = tuple(config for config in self._configs if config.split != "test" or config.year == "2007")
  	def make_config(self, **options: Any) -> DatasetConfig:
  	config = super().make_config(**options)
  	if config.split == "test" and config.year != "2007":
  	raise ValueError("`split='test'` is only available for `year='2007'`")
  	return config

  with

  vision/torchvision/prototype/datasets/_builtin/voc.py

  Lines 54 to 55 in 3be12c7

  	if split == "test" and year != "2007":
  	raise ValueError("`split='test'` is only available for `year='2007'`")

• Instead of passing a config object around, the dataset can now simply use instance attributes set in the constructor. Compare

  vision/torchvision/prototype/datasets/_builtin/cub200.py

  Lines 93 to 98 in 5f74f03

  	def _2011_filter_split(self, row: List[str], *, split: str) -> bool:
  	_, split_id = row
  	return {
  	"0": "test",
  	"1": "train",
  	}[split_id] == split

  
  

  vision/torchvision/prototype/datasets/_builtin/cub200.py

  Line 174 in 5f74f03

  split_dp = Filter(split_dp, functools.partial(self._2011_filter_split, split=config.split))

  with

  vision/torchvision/prototype/datasets/_builtin/cub200.py

  Lines 116 to 121 in 4c73a5e

  	def _2011_filter_split(self, row: List[str]) -> bool:
  	_, split_id = row
  	return {
  	"0": "test",
  	"1": "train",
  	}[split_id] == self._split

  
  

  vision/torchvision/prototype/datasets/_builtin/cub200.py

  Line 192 in 4c73a5e

  split_dp = Filter(split_dp, self._2011_filter_split)

• Although not yet added, we can now document the dataset with a regular docstring. This was not possible before, because the old Dataset object would have not been exposed and in turn would not show up in the documentation.

• We can use the native __len__ method to return the number of samples in the dataset. In the old implementation we would have needed to attach a custom datapipe that does nothing but adds this annotation.

[pmeier]

There are a few open PRs that need to change when this is merged:

• Port SBU dataset #5683
• Add places365 datapipe #5679
• Add sun397 prototype datapipe #5667
• add WIDERFace prototype dataset #5651
• Implementation for cityscapes in proto datasets #5441
• add LFWPeople prototype dataset #5438
• add HMDB51 and UCF101 datasets as well as prototype for new style video decoding #5422
• add LSUN prototype dataset #5390
• Add Sintel Dataset to the dataset prototype API #4895

A couple of them are from OSS contributors. Given that this rework only shuffles some information around, I propose I'll push some changes to their branches to minimize the impact.

In addition, after this is merged, I'll post an issue in #5336 for everyone who already started but hasn't send a PR yet.

Finally, with this change, are we now confident enough to re-open the dataset ports to community contributions?

[NicolasHug]

Thanks a lot @pmeier

The prototype job is green so it's safe to approve now, but to be extra careful let's wait until all jobs are green before merging.

The summary above is very useful, thank you. On top of what you already said, it's worth noting that even though this isn't reflected in this PR, the number of line of code has significantly reduced in our testing + base data structures, by ~500 (this PR reflects this more accurately: #5774). The reason why this PR shows more loc is because we added more features (e.g. support for __len__()), and more docs. Thanks again for your help Philip

[NicolasHug]

Also CCing @ejguan @NivekT @VitalyFedyunin ; Thanks to you guys' work on fixing pytorch/data#237, we were able to greatly simplify our implementations, as described above. As you know we still need a full-fledged fix that also handles environments where dill is installed, but for now we're able to move forward with our preferred design. Thank you!