Move image loading logic to _load_image in VOC datasets (similar to COCO)

[vadimkantorov]

🚀 The feature

This would allow simple stubbing out image loading if needed for fast dataset transformed metadata access (when image reading is not needed) by dataset._load_image = lambda _: None and also make API more consistent across datasets.

It's also impossible to monkey-patch __getitem__ at instance level: https://stackoverflow.com/questions/57413453/is-it-possible-to-override-getitem-at-instance-level-in-python :(

Motivation, pitch

N/A

Alternatives

No response

Additional context

No response

[NicolasHug]

I would recommend to inherit from the class, and override __getitem__ instead then, it should only be a few more lines of code.
If you submit a PR for this I think we would merge it, but keep in mind the _load_image function is private API, so while I know you know what you're doing, we don't offer any guarantee w.r.t. BC or consistency across datasets here.

[vadimkantorov]

Here's what I do now: dataset.__getitem__ = lambda index: dataset.transforms(None, dataset.parse_voc_xml(torchvision.datasets.voc.ET_parse(dataset.annotations[index]).getroot()))

Overriding getitem would entail to the same thing - having to copy-paste the even less standard / error-prone annotation parsing line.

Not wanting the dataset to load images for only metadata purposes is a valid and frequent usecase (often happens for some stats computation and avoiding duplicating ground truth preparation functions). At least simplifying hacking around would be helpful and minimize the copy-pasting. Accepting a method for image loading by a given path in __init__ is another way (probably for pickle purposes, there would need to be a torchvision method function None)

[pmeier]

As discussed in #4991, I'm ok with moving the functionality into a private method on the dataset as long as it is a 1-to-1 copy of the current behavior. Still, the BC part of #4975 (comment) applies.