Hi Daniel! Some precisions:
I think I wasn’t clear enough: multinomial won’t sample (at least not out of the box, you can adapt it if needed) if you have some categories with a probability of 0, but I don’t think it’s your case. Categorical will of course sample, as a Categorical is just a generalization of Bernoulli to more than two categories.
Did you do it with pandas? It should be straightforward: df.astype(int). More generally, are you doing your data cleaning in pandas? Because it’s weird that you’re doing it with the tensors – i.e with theano.