Rework train-val-test split

[florinbarbisch]

• [x] use 6 instead of 5 folds
• [x] use a separate fold for testing (don't use the validation fold for this)
  • [x] stats per fold
  • [x] Dataset param: fold_indicies (tuple, list)
  • [x] stats per fold aggregieren
• [x] Create stratified folds
  • [x] use the following heuristic:
  • [x] create a metric that describes the equality of the distribution: a talk with chatGPT about possible metrics and the prevention of overfitting
  • [x] init each fold with one chip
  • [x] for each chip add it to each fold and compute the new metric
  • [x] add the chip to the fold that created the best score
  • [x] repeat until all chips are distributed
  • [ ] ~additions~
    • [x] add a penalty against an unbalanced number of chips in each fold
    • [x] use log distribution instead of normal pixel count
    • [ ] add a penalty against an unbalanced number of pixels (without background) in each fold
    • [ ] add more regularization techniques against overfitting
  • [x] for speed up: aggregate stats for each fold and just +/- the stats for the to be added/removed chip