Could T.Lambda be nn.Module?

[vadimkantorov]

It would allow it to be placed in nn.ModuleList for passing to RandomApply (for scriptability)

https://pytorch.org/vision/stable/transforms.html#torchvision.transforms.RandomApply

cc @vfdev-5

[fmassa]

Hi,

@vfdev-5 has more context on this and please correct me if I'm missing something.

We tried doing it, but it wouldn't be possible to make it scriptable IIRC, and there were some other things that could be misleading to users.

There is some more discussion in #2645, in particular #2645 (comment), but I'll let @vfdev-5 give more details

[vadimkantorov]

Real-life usecase: facebookresearch/dino#28 (also partly because Solarization does not exist as a module)

[fmassa]

@datumbox what do you think about adding a nn.Module interface for Solarize?

[datumbox]

We already support this, unless you mean something else?

vision/torchvision/transforms/transforms.py

Lines 1828 to 1837 in 2f3b7ca

	class RandomSolarize(torch.nn.Module):
	"""Solarize the image randomly with a given probability by inverting all pixel
	values above a threshold. If img is a Tensor, it is expected to be in [..., 1 or 3, H, W] format,
	where ... means it can have an arbitrary number of leading dimensions.
	If img is PIL Image, it is expected to be in mode "L" or "RGB".
	Args:
	threshold (float): all pixels equal or above this value are inverted.
	p (float): probability of the image being color inverted. Default value is 0.5
	"""

[vadimkantorov]

Unfortunately, it's missing from the docs: http://pytorch.org/vision/master/search.html?q=RandomSolarize&check_keywords=yes&area=default <- no results

[datumbox]

@vadimkantorov Thanks for flagging. We really appreciate taking the time to do so. Keep them coming. :)

I just sent a PR to fix this. Should be added shortly.