Defined hyperspectral-specific Transforms

[nshaud]

torchvision defines Transforms objects to apply data augmentation and other transformations to data.

We could and should define our own custom Transforms.

Pros:

• easier to reuse the data augmentation transforms that we implement
• easier to apply various kinds of DA in the toolbox

Cons:

• adds an external dependency to torchvision

[mengxue-rs]

@nshaud Sorry bother you, hyperspectral-specific Transforms should include what kinds of transforms? do you mean dimension reduction methods like principle component transform and maximum noise fraction or data augmentation methods like radiation-based virtual samples and mixture-based virtual samples?

[nshaud]

Yes, I was mainly thinking data augmentation. Other augmentations such as dimension reduction could come later but I don't see any benefit to adding them now.