Normalizing pixel values

[roybenhayun]

in many cases, normalizing input data to a range between 0 and 1 can improve the performance of neural networks.
how and what typically would be a good normalization of images and masks in torchgeo?

for example, when using Sentinel2, the pixels values are 0-10000 (i.e, see how Sentinel2.plot() "normalizes" and clamps the values for local rendering purpose).
so should a transform, which does a similar normalization, be passed to the Sentinel2 dataset c'tor?
s2_ds = Sentinel2(root=sn2_root, bands=Sentinel2.rgb_bands, transforms=preprocess)
however, this will break the Sentinel2.plot()...

so the questions are:

• should Sentinel2 or other datasets values, be normalized
• if yes, to which range, and how.
• also, how to ensure such normalization does not break plotting

thanks!

[adamjstewart]

should Sentinel2 or other datasets values, be normalized

Yes, you almost always want normalization.

if yes, to which range, and how.

This is up to the user. Popular choices include mapping to the 0–1 range or subtracting the mean and dividing by the standard deviation. I would personally recommend using Kornia to apply normalization to an entire batch of data at a time instead of passing a transform to the dataset. This is what we do in our data modules.

also, how to ensure such normalization does not break plotting

See #1263 for a proposal of a better way to handle this.