Sander Dieleman

Seems a bit arrogant maybe to hijack 'recipes' for this? I don't know :) It would be the most convenient option I suppose.

Yeah I don't suppose this is really PyPI material. Especially because it's largely untested and undocumented.

I don't know, it might be confusing if half of the repository content is in the top-level and half of it is in the `recipes` subdirectory. Also we have to...

I don't have much time to work on this right now (beyond cosmetic improvements) so I'm okay with moving it to 'examples' as well if you prefer.

I was actually planning to have a look at it today. But I can do a new PR after you merge them as well, up to you.

Sure :) I'll try to sort out @f0k's remarks by tonight.

Just tried initializing the biases to `-2.0` instead of `-4.0`, but this severely slows down convergence and seems to make things slightly unstable as well. So I'll stick with `-4.0`....

I tried changing some parameter values to match the paper better, but in the end I decided to leave most of them as they were because it just caused trouble...

Yeah, I think we should also keep them here, this is where people will come looking for them. A separate subdirectory for pre-trained models is not a bad idea.

there's no guarantee that this (remarkably well working) analogy will keep being productive though :p