Mod add seed issue

[nix-apollo]

From @danbraunai-apollo:

This is also an issue/bug for ModularArithmetic. We have an even worse setup there with this code in create_modular_arithmetic_dataset:

    modulus = cfg["dataset"]["modulus"]
    fn_name = cfg["dataset"]["fn_name"]
    frac_train = cfg["dataset"]["frac_train"]
    seed = cfg["seed"]

modulus = dataset_config.modulus or modulus
fn_name = dataset_config.fn_name or fn_name
frac_train = dataset_config.frac_train if dataset_config.frac_train is not None else frac_train
seed = dataset_config.seed if dataset_config.seed is not None else seed

This looks for a main config seed in the saved model (as opposed to looking for seed in the dataset config), and replaces it with a seed in the dataset config.

[nix-apollo]

I'm confused by what the problem is, @danbraunai-apollo should clarify.

[danbraunai-apollo]

I think the code above should have the line

    + dataset_seed = cfg["dataset"]["seed"]
    - seed = cfg["seed"]

and then passing dataset_seed instead of seed where used in the future.

Also, since this value can be None, we have to handle passing seed=None to train_test_split in create_modular_arithmetic_dataset. The best thing to do IMO is to have a thing at the start of all modular arithmetic scripts which does config.dataset.seed = config.seed if config.dataset.seed is None.

The way the code current works is "let's use config.dataset.seed to manage the dataset when training the model. But if loading a trained model, let's use config.seed."