Greyscale image generation about lightweight-gan

[lucidrains]

https://github.com/lucidrains/lightweight-gan/releases/tag/0.14.0 ok done with the --greyscale flag

[stefanorosss]

You are definitely right! I started with the assumption of wanting to change from RGB and didn't question it later, thanks

[woctezuma]

Just a quick comment: wouldn't it have been simpler to edit the logic the other way around? Rather than keeping "transparent" and replacing RGB with grayscale, you could have kept RGB and replaced "transparent" (4 channels) with "grayscale" (1 channel). In this case, all you would have to do is Ctrl+F "transparent" (32 matches) in https://github.com/lucidrains/lightweight-gan/blob/main/lightweight_gan/lightweight_gan.py and change the logic every time.

Here, you have the expand_greyscale class, which risks setting the number of channels back to 3:

lightweight-gan/lightweight_gan/lightweight_gan.py

Lines 184 to 208 in 15a1295

	class expand_greyscale(object):
	def __init__(self, transparent):
	self.transparent = transparent
	def __call__(self, tensor):
	channels = tensor.shape[0]
	num_target_channels = 4 if self.transparent else 3
	if channels == num_target_channels:
	return tensor
	alpha = None
	if channels == 1:
	color = tensor.expand(3, -1, -1)
	elif channels == 2:
	color = tensor[:1].expand(3, -1, -1)
	alpha = tensor[1:]
	else:
	raise Exception(f'image with invalid number of channels given {channels}')
	if not exists(alpha) and self.transparent:
	alpha = torch.ones(1, *tensor.shape[1:], device=tensor.device)
	return color if not self.transparent else torch.cat((color, alpha))

which is then called at the end here by transforms.Lambda(expand_greyscale(transparent)):

lightweight-gan/lightweight_gan/lightweight_gan.py

Lines 222 to 232 in 15a1295

	convert_image_fn = partial(convert_image_to, 'RGBA' if transparent else 'RGB')
	num_channels = 3 if not transparent else 4
	self.transform = transforms.Compose([
	transforms.Lambda(convert_image_fn),
	transforms.Lambda(partial(resize_to_minimum_size, image_size)),
	transforms.Resize(image_size),
	RandomApply(aug_prob, transforms.RandomResizedCrop(image_size, scale=(0.5, 1.0), ratio=(0.98, 1.02)), transforms.CenterCrop(image_size)),
	transforms.ToTensor(),
	transforms.Lambda(expand_greyscale(transparent))
	])

Apart from that, there are the changes below which you have already mentioned:

• generator:
  

  lightweight-gan/lightweight_gan/lightweight_gan.py

  Line 338 in 15a1295

  init_channel = 4 if transparent else 3

• discriminator:
  

  lightweight-gan/lightweight_gan/lightweight_gan.py

  Line 484 in 15a1295

  init_channel = 4 if transparent else 3

• image export:
  

  lightweight-gan/lightweight_gan/lightweight_gan.py

  Lines 1170 to 1172 in 15a1295

  	if self.transparent:
  	background = Image.new('RGBA', pil_image.size, (255, 255, 255))
  	pil_image = Image.alpha_composite(background, pil_image)

[bogdan-ivan]

This would be great to be added as an option. @lucidrains

[lucidrains]

@bogdan-ivan sure, i'll add it! what is the corresponding enum string in pillow for greyscale? (like RGBA is for transparent)

[woctezuma]

@bogdan-ivan sure, i'll add it! what is the corresponding enum string in pillow for greyscale? (like RGBA is for transparent)

The last time I checked, it was L.

https://pillow.readthedocs.io/en/stable/handbook/concepts.html#modes

[Dok11]

@lucidrains I think about --num-channels flag. For example to generation textures which contain several maps (albedo, roughness, specular, normal map). How do you think does possible modify code to work with them? And will it be too hard or easy?