seed for the random numbers generators

[sdv-dev]

seed for the random numbers generators

[ManuelAlvarezC]

Hi @echo66

I agree that being able to set the random_state to ensure reproducibility of your results is indeed a useful feature to have.

From the approaches you propose, I'll definetely pick the second one: having random_state as an constructor argument on the classes that use some kind of stochastic process.

I would also take your suggestion of using only one of random and numpy.random. Will you be open to do the changes yourself?

[ManuelAlvarezC]

The required changes should be the following:

1. Replace all the ocurrences of random with np.random
2. Create a method random_state_call on copulas.__init__, like this:

def random_state_call(random_state, function, *args, **kwargs):
    if random_state is None:
        return function(*args, **kwargs)

    else:
        original_state = np.random.get_state()
        np.random.seed(random_state)
        
        result = function(*args, **kwargs)
        
        np.random.set_state(original_state)
        return result

3. Change classes bivariate.base.Bivariate and multivariate.tree.Tree __new__ method in order for them to accept arbitrary positional and keyword arguments *args, **kwargs

4. Add the argument random_state=None on __init__ for both classes, that sets its value to the attribute of the same name.

5. Replace on their subclasses all calls to functions that depend on their np.random with calls to random_state_call, that is:

np.random.uniform(0, 1, n_samples)

random_state_call(self.random_state, np.random.uniform, 0, 1, n_samples)

6. Repeat the steps 4 & 5 for univariate.base.Univariate, multivariate.base.Multivariate and multivariate.vine.VineCopulas.

7. Create unittests for random_state_call and for the sample methods, in which, after a random state is set, the sample values are exactly the same.

8. Check everything is according to our Contributing Guide.

If you plan on updating your PR with this changes, please let us know and we will assign this issue to you.