Add option to seed random number generator

[cbrnr]

In #89, we added a random seed to our examples. I think it would be nice if we had an option to set the seed globally, e.g. ws.set_seed for the OO API.

[cbrnr]

We could add a seed option to every function/method that uses random numbers.

[mbillingr]

I don't like messing with the numpy global seed inside our functions. It's probably better to use an explicit instance of random state that can be passed to functions instead of a seed. (Default is to use global state)

[cbrnr]

scikit-learn functions have a random_state argument: http://scikit-learn.org/stable/modules/generated/sklearn.cross_validation.KFold.html#sklearn.cross_validation.KFold

def check_random_state(seed):
    """Turn seed into a np.random.RandomState instance

    If seed is None, return the RandomState singleton used by np.random.
    If seed is an int, return a new RandomState instance seeded with seed.
    If seed is already a RandomState instance, return it.
    Otherwise raise ValueError.
    """
    if seed is None or seed is np.random:
        return np.random.mtrand._rand
    if isinstance(seed, (numbers.Integral, np.integer)):
        return np.random.RandomState(seed)
    if isinstance(seed, np.random.RandomState):
        return seed
    raise ValueError('%r cannot be used to seed a numpy.random.RandomState'
                     ' instance' % seed)

[mbillingr]

That was exactly what I had in mind.

The Workspace can have its own RandomState that is passed to all functions. All other functions/classes take either a RandomState argument, a seed for constructing a new RandomState, or they default to using numpy's global state.

This is even perfectly backwardscompatible.

[cbrnr]

But do you think the Workspace can set the random state, i.e. ws.seed(123)? This would then set the Numpy seed.

[mbillingr]

I think it's bad practice to implicitly change the behavior of a completely different package.

Consider this:

ws.seed(123)  # calls np.random.seed(123)
x = np.random.randn()  # this is affected by the seed

If we do a ws.seed(123) everything outside the workspace should be unaffected.

[cbrnr]

OK. But this means we can still have a ws.seed(123), which sets the seed within the workspace (which is passed on to all methods and functions).

[mbillingr]

Yes. Well, not the seed, but the RandomState object.

[cbrnr]

OK, got it! Now which functions need a random_state argument?

[mbillingr]

I think only these: Whiteness test, connectivity statistics, VAR simulation.

[cbrnr]

I think all ICA functions too.

[mbillingr]

True. We have to check our modification of Infomax. Maybe I edited out similar functionality.