draw_sample implementation

[lbollar]

Hello team, I hope you don't mind educational questions here. I saw Snorkel had a lot of these so I thought you didn't.

I have been diving into the Snorkel papers and code and that led me to Numbskull. I had read in the papers that Gibbs sampling was used and I was a little confused in the papers what was actually being sampled.

On this line here, it looks like you are taking one of the exponentiated potentials and multiplying by a uniform random sample which is then used in some argmax logic. I was wondering what was going on here with this random uniform sample and how it is implementing a gibbs sampling.

If this is annoying feel free to close :)