set_seed in train_model

[prajjwal1]

Patch description
Adding a function that allows user to set seed for training. User just has to pass the flag --set_seed with existing arguments to train_model.py in case they want to use a seed other than 42.

Testing steps

Other information

[prajjwal1]

@stephenroller Can you please look into this ? It has been here for sometime now.

[prajjwal1]

I was thinking of using the functionality brought from this to teachers also. Teacher files have SEED manually hardcoded to 42. How is a user supposed to set the seed in Teacher. This will allow user to bring this functionality to control rng in teacher.

[stephenroller]

I'm fine with this, but I'm pretty sure the existing code already sets the seed. Can we double check?

[prajjwal1]

I did check, i didn't find any way to setting the seed universally. I plan to send one more PR for multiwoz_v22, which will set the rng through opt. Current opt dict doesn't have it, so where else is seed parameter supposed to go ?

[klshuster]

I'm fine with this, but I'm pretty sure the existing code already sets the seed. Can we double check?

interestingly I can't seem to find it either. we set seeds in nearly every other script except for train model

pure non-determinism! In that case, can we have the default seed be None and only set seed if it's not None?

[prajjwal1]

Default way is to set the seed to 42 right ? Can I set it to 42, and this can be overridden by the user by passing in --seed parameter ? By not setting the seed, users are likely to observe variability in their training runs.

[stephenroller]

I'm fine with this, but I'm pretty sure the existing code already sets the seed. Can we double check?

interestingly I can't seem to find it either. we set seeds in nearly every other script except for train model

pure non-determinism! In that case, can we have the default seed be None and only set seed if it's not None?

ah right this is what i misremembered

[stephenroller]

Default way is to set the seed to 42 right ? Can I set it to 42, and this can be overridden by the user by passing in --seed parameter ? By not setting the seed, users are likely to observe variability in their training runs.

historically we've preferred backwards compatibility

[klshuster]

I agree, let's keep it to None. We've advocated for variability from the start

[prajjwal1]

@klshuster Is it fine now ?

[stephenroller]

lgtm, will leave Kurt for final approval and merge

[klshuster]

thank you!