Randomness & Reproducibility

[dilyabareeva]

• track the source of randomness
• make sure that results are as reproducibly as possible

[gumityolcu]

https://numpy.org/doc/stable/reference/random/generator.html

Using dedicated RNG will solve this 👍🏼 With correct versions of course. I seem to remember an equivalent class in torch. What I can find is this
https://pytorch.org/docs/stable/generated/torch.Generator.html

But it doesn't seem to be exactly the same thing-.

[gumityolcu]

@dilyabareeva my proposal is given in the randomizartion metric. takes a seed from the user and initiates a torch.Generator object to use for randomization. This way, using the same exact version of torch with the same seed will ALWAYS give the same random numbers (since no OS based randomizer is used, everything is done with python inside torch)

[dilyabareeva]

Lightning provide a "reproducibility" option for model training: https://lightning.ai/docs/pytorch/stable/common/trainer.html#reproducibility