resize input to fill input tensor

[claudius-kienle]

Currently, the inputs are not resized at all.
This results in large inputs with just a few features.
It would make sence to resize the input to fill the whole input tensor.

Things to keep in mind:

• Should the resizing be the same on the whole dataset, or should it be computed pixel wise?
• It should be possible to undo the resization after inference