Julia Evans

Day 10: Training an RNN to count to three

Yesterday I was trying to train an RNN to generate English that sounds kind of like Shakespeare. That was not working, so today I instead tried to do something MUCH simpler: train an RNN to generate sequences like

0 1 2 0 1 2 0 1 2 0 1 2

and slightly more complicated sequences like

0 1 2 1 0 1 2 1 0 1 2 1 0 1 2 1 0

I used (I think) the exact same RNN that I couldn’t get to work yesterday to generate English by training it on Shakespeare, so it was cool to see that I could at least use it for this much simpler task (memorize short sequences of numbers).

the jupyter notebook

It’s late so I won’t explain all the code in this blog post, but here’s the PyTorch code I wrote to train the RNN to count to three.

In the gist there are a few experiments with different sequence lengths, like (unsurprisingly) it takes longer to train it to memorize a sequence of length 20 than a sequence of length 5.

simplifying is nice

I’m super happy that I got an RNN to do something that I actually understand! I feel pretty hopeful that on Monday I’ll be able to go back to the character RNN problem of trying to get the RNN to generate English words now that I have this simpler thing working.

Day 9: Generating a lot of nonsense with an RNN New zine: Hell Yes! CSS!