Discussion about default unnormalize method for pytorch tensor

[genghisun]

Currently, it seems that for input pytorch float tensor between -1 and 1, a min-max normalization is taken:

torchshow/torchshow/visualization.py

Lines 264 to 287 in 3008c23

	def auto_unnormalize_image(x):
	all_int = isinteger(np.unique(x)).all()
	range_0_1 = within_0_1(x)
	range_0_255 = within_0_255(x)
	has_negative = (x.min() < 0)
	if has_negative:
	logger.debug('Detects input has negative values, auto rescaling input to 0-1.')
	return rescale_0_1(x)
	if range_0_255 and all_int and (not range_0_1): # if image is all integer and between 0 - 255. Normalize it to 0-1.
	logger.debug('Detects input are all integers within range 0-255. Divided all values by 255.')
	return x / 255.
	if range_0_1:
	logger.debug('Inputs already within 0-1, no unnormalization is performed.')
	return x
	logger.debug('Auto rescaling input to 0-1.')
	return rescale_0_1(x)
	def rescale_0_1(x):
	"""
	Rescaling tensor to 0-1 using min-max normalization
	"""
	return (x - x.min()) / (x.max() - x.min())

However, usually people use transforms.Normalize((0.5,), (0.5,)) to normalize image.
So I think when x is between -1 and 1, it should be x * 0.5 + 0.5 to keep the image color consistent.

I know it can be achieved by setting set_image_mean([0.5, 0.5, 0.5]) and set_image_std([0.5, 0.5, 0.5]), but this may be used for other special cases. For the most common cases, i think we should adopt the above method.

Any idea?

[xwying]

Good suggestion! I think we could do it if this is a common normalization preset. In fact, what you described can be extended to all the zero-centralize normalizations as well. The tricky part that I can think of is how to determine if the data is zero-centralized rather than just happen to fall between that range. In other word, is simply checking the [-1, 1] range (or [-M, M] for more general cases) sufficient to determine these kind of inputs?

[genghisun]

I think simply checking [-1, 1] is enough. Is there any other possible situation? For normal images it seems that these are the only cases. For other special cases, they can use set_image_mean and set_image_std.

[xwying]

Right, maybe we can try to handle the [-1, 1] case for now. Would you be able to create an PR for this feature and some testing cases so we can review and discuss further?