deterministic permutation function

[jettjaniak]

• context
  • we want reproducible training
  • one of the elements of that is ensuring model is seeing samples in the same order in every run
  • we need to shuffle samples in each epoch
• we will go with a deterministic function in Python
  • permute(input_output: list, seed: int) -> None
    • it returns nothing, but permutes the list in-place
    • permutation doesn't depend on the elements
    • but only on the seed and length of the input
    • put it in train/permute.py
• we don't want to use any libraries, as RNG engines in these can change over time
• here's a (buggy) idea how to do it https://chat.openai.com/c/b4c2a457-1fb3-4f20-8016-6bdf71a0a2d3
  • use some reasonable LCG parameters (ideally from an old classic implementation)