Remove color_space metadata and ConvertColorSpace() transform

[NicolasHug]

As title.

The rationale for removing the color_space metadata is that there's a 1:1 mapping between the color-space and num_channels for all currently supported color-spaces. So color_space metadata is redundant. There's no plan to add support for additional color-spaces at this time (there could be in the future, at which time we could re-consider such meta-data).

The rationale for removing the ConvertColorSpace() transforms / functionals is that there's no obvious current need for those. We don't have such transforms in the current transforms. Obtaining an image with the desired color-space can currently be achieved by pil_to_tensor(Image.Open(...).convert("RGB")) for PIL images or read_image(..., mode=ImageReadMode.RGB) for tensors. (Also note the unfortunate overalp between the ColorSpace and ImageReadMode enums - something to consider in the future!)

---

Right now I'm being lazy and I moved the ColorSpace enum into the tests. The changes look big but it's mostly just s/datapoints.ColorSpace/ColorSpace/g.
However ideally, we probably want to get rid of ColorSpace completely and rely on num_channels instead. Waiting on @pmeier feedback on this first, as this seem to require pervasive changes throughout the tests which I'm not too familiar with yet.

Still TODO in another PR:

• un-deprecate the [Random]GrayScale() transforms and make them rely on _rgb_to_gray().

cc @bjuncek @pmeier

[pmeier]

Thanks Nicolas! Tests are failing because rgb_to_grayscale uses a function to determine the color space that was removed by you.

vision/torchvision/prototype/transforms/functional/_deprecated.py

Line 31 in d2d448c

old_color_space = datapoints._image._from_tensor_shape(inpt.shape) # type: ignore[arg-type]

Since that is only used to construct the warning, we can just basically scrap it all together. I'll leave it up to you to actually un-deprecate here or in a follow-up PR.

[pmeier]

This can probably go, since it emulates the metadata on the actual datapoint.

[pmeier]

Yup, that seems like an artifact. Basically the cls._to_tensor() and super().__new__ do the same thing. Let's remove the latter since it is obsolete.

[pmeier]

Yeah, let's do that in a follow-up.

[pmeier]

IIUC, the difference is that to_grayscale is just for PIL images and thus supports arbitrary input color spaces or rather all that Pillow supports. Although there is no functional difference for PIL images in rgb_to_grayscale, tensors have to have the three RGB channels.

+1 on deprecating.

[pmeier]

However ideally, we probably want to get rid of ColorSpace completely and rely on num_channels instead. Waiting on @pmeier feedback on this first, as this seem to require pervasive changes throughout the tests which I'm not too familiar with yet.

I agree, we shouldn't have enums in the test suite that don't exist in the library. Just to clarify: do you want to get rid of the enum or the general functionality? IMO, make_image(..., color_space="RGB") is quite a bit more expressive than make_image(..., num_channels=3).

[pmeier]

Lint job is failing: https://app.circleci.com/jobs/github/pytorch/vision/1810524. Otherwise LGTM, thanks Nicolas!

[github-actions]

Hey @NicolasHug!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py

[vadimkantorov]

The rationale for removing the ConvertColorSpace() transforms / functionals is that there's no obvious current need for those.

colorspace conversion functions are useful in contexts other than augmentations, related discussion: #4029 and such functions are present in most computer vision libraries