Controlling randomness for reproducibility

The current implementation of ivadomed only sets the random number generator seed for splitting a dataset into train, validation and test sets. However, there are other sources of randomness that we need to take into consideration, such as 1) the Numpy library 2) Pytorch library 3) Python itself 4) CUDA etc. We may use this reference

The main reason is seeding the dataset splitting alone, still leaves some sources of randomness unchecked. Also, if a user decides to use a .joblib file to split the data as documented here , the random_state won’t be used, but we still need to control other sources of randomness.

[jcohenadad]

@etetteh can you please open a PR so we can follow your development

[uzaymacar]

Hey @etetteh I have a couple thoughts / suggestions but let me know if I’m missing something:

• You want to do context[ConfigKW.RANDOM_SEED] but context is not defined in ivadomed/loader/utils which is the current call place for set_seed
• Instead of calling set_seed here we probably want to call inside main.py
• run_command includes context as an input parameter so that might be one place to call set_seed. This is also good for the visibility and logical ordering of the function and as we usually want to set seeds the first thing in the script.