Value range after applying ToPILImage()

[sahilg06]

📚 The doc issue

Here it is written that ToPILImage() preserves the value range, Doesn't it multiply by 255. as done here?

Suggest a potential alternative/fix

No response

[pmeier]

You are right, the wording is confusing. What it means to convey is that we automatically adjust the value range for the user if needs be. Meaning, if you omit mode for a three channel image, we assume

vision/torchvision/transforms/transforms.py

Line 213 in 8324c48

- If the input has 3 channels, the ``mode`` is assumed to be ``RGB``.

And since there is no floating point RGB support in PIL, we need to convert the value range to do the thing the user wants, i.e. get a PIL image that is as close as possible as the original tensor.

Note that if you have a grayscale tensor image, i.e. either of shape (1, H, W) or just (H, W), and just explicitly set mode="F", we don't perform a value range conversion

vision/torchvision/transforms/functional.py

Line 290 in 8324c48

if pic.is_floating_point() and mode != "F":

image_tensor = torch.rand(256, 256)
image_pil = ToPILImage(mode="F")(image_tensor)

torch.testing.assert_close(np.array(image_pil), image_tensor.numpy(), atol=0, rtol=0)

We would be happy to receive a PR clarifying that.

[pmeier]

Also, we want to have a proper listing in in the mode documentation

vision/torchvision/transforms/transforms.py

Lines 211 to 216 in 8324c48

	If ``mode`` is ``None`` (default) there are some assumptions made about the input data:
	- If the input has 4 channels, the ``mode`` is assumed to be ``RGBA``.
	- If the input has 3 channels, the ``mode`` is assumed to be ``RGB``.
	- If the input has 2 channels, the ``mode`` is assumed to be ``LA``.
	- If the input has 1 channel, the ``mode`` is determined by the data type (i.e ``int``, ``float``,
	``short``).

but it renders as regular text. We probably need a blank line after the colon.

[sahilg06]

Hi @pmeier, I would be happy to create a PR for the same.
Thank you!