Variables in configs not user-friendly

[fabioseel]

Still not sure how to feel about referencing those variables here and thus forcing the experiment config to set them. This is somewhat obscure and non intuitive.
However, I'm also not fundamentally happy about the alternative (just have defaults here and overwrite them in the experiment config) - even though it's more explicit and imho easier to use at first sight, intrinsic dependencies (eg vision_height and _width definition between model and dataset needs to match) can easily be messed up (but this also applies for the current system.

In general, I think this is user unfriendly and we should rediscuss and fix this somehow.

[alex404]

Yah I have thoughts about this too but its not urgent. We can discuss next time it makes sense.