Image/label values normalisation

Issue description

Sometimes, the data from nifti or from h5 might not have the range as we expected.

For now, we expect:

• Images have values between 0-255
• Labels have values between 0-1

And so we will divide the image tensors by 255.

But sometimes this might not be the case and maybe user would like to define their own value range. So we can do the following instead,

• For each file loader, h5 or nifti, they will need a value range [min, max] and they will normalize the arrays using the given values by x = (x - min) / (max - min + EPS) with EPS=1.0E-6 a very small number. The min and max can be None (any one of them, not necessarily both), and if None, the min/max will be calculated dynamically on the fly. So the min/max would not be global, but be array dependant.
• Then, for image file loader we will assume the max = 255, min = 0, for label file loader we will assume the max = 1, min = 0. However, user can overwrite this values via config. But we do not have to put these values in config, it’s optional. Otherwise the config will be too complicated.

What do you think? @ebonmati @NMontanaBrown @YipengHu

Type of Issue

• New feature request
• Documentation update
• Test request

[YipengHu]

I divide it by 255 because I don’t want to feed it with different ranges. So fine, agree that with batch norm the difference is little. Will change so, and change the configs to make it clear, unless given min/max in the config, we will normalize dynamically.

@mathpluscode i see your point - just conventions - pick one and clarify it in config/doc.

[YipengHu]

It is error-prone with anything “we assume”. i think by default, it should be x = (x - min) / (max - min + EPS) - calculated per image, unless a global customised [min max] is given; - no need for 255 as a special case.

Another important normalisation is unit-variance with zero-mean. Not sure i see clear advantage, so let’s leave it post-r1 when there is a need.