Random seeds

[kwalcock]

Instead of this code

np.random.seed(i)
np.random.shuffle(trainX)
np.random.seed(i)           # make sure to reset seed again to keep labels and data together!
np.random.shuffle(trainY)

you might try something more like

rng_state = np.random.get_state()
np.random.shuffle(trainX)
np.random.set_state(rng_state)
np.random.shuffle(trainY)

It's unfortunate that the numpy shuffle routine doesn't take a random generator as an argument so that two different generators can be used in parallel. The new code rewinds the single generator to the old state between sorts. This might be more random than sampling the initial values of consecutive sequences and it won't mess up randomization elsewhere in the code if the reseed is not expected.

[michaelcapizzi]

That's a good idea @kwalcock . Generally, in my NLP work, I'm working with <list>s of text which are easy to shuffle using shuffle(). But in this case, having a numpy array was something I hadn't had to deal with before.

Thanks for the suggestion.

If you want to make a pull request, I'll happily merge this into the notebooks.