Add support for VOC on detection references

[pmeier]

The segmentation references already support VOC and COCO

vision/references/segmentation/train.py

Lines 20 to 24 in 08c8f0e

	paths = {
	"voc": (dir_path, torchvision.datasets.VOCSegmentation, 21),
	"voc_aug": (dir_path, sbd, 21),
	"coco": (dir_path, get_coco, 21),
	}

Unfortunately, the detection references only support COCO

vision/references/detection/train.py

Line 37 in a1232c2

paths = {"coco": (data_path, get_coco, 91), "coco_kp": (data_path, get_coco_kp, 2)}

There was a first push for this in #1216 by @lpuglia. We can probably salvage some parts from the PR.

We are currently reworking the datasets and transforms, and in turn the reference scripts will change quite a bit. Thus, we should only go for this after they are stable-ish again.

cc @datumbox @YosuaMichael

[ambujpawar]

This seems to be in interesting.
Is this still valid? Can I pick this up?
cc @pmeier

[pmeier]

Thanks for the interest in this @ambujpawar. Quoting myself from above:

We are currently reworking the datasets and transforms, and in turn the reference scripts will change quite a bit. Thus, we should only go for this after they are stable-ish again.

Unfortunately, we haven't reached a stable design yet. Transforms are about 80-90% done and there is #6433 to integrate them into our references. However, for datasets we are not yet ready to put them into an end-to-end tests like our references. Thus, going for this now is probably wasted time.

If you are still interested, I can ping you if the references have stabilized and you can go for it then. How does that sound?

[ambujpawar]

Thanks for the quick response.
Yeah, that sounds perfect!!

In the meantime, I'll find other issues to contribute :)