Update: You can see a web app in action here.

Like the rest of America, I’ve spent a lot of time the past several days playing Pokemon Go. I also spent some time learning some basic things about Recurrent Neural Networks (RNNs) for generating random text. Previously, I had used simpler methods for generating random text.

But since “Deep Learning” sounds cool (and is cool), I decided to use this sophisticated method to generate random Pokemon names! There are already some web projects that create Pokemon mashups and can generate random teams of Pokemon or random Pokemon names. But I didn’t see anything that did precisely what I wanted–the closest was Andrej Karpathy’s baby names.

I set up Justin Johnson’s torch-rnn project (based on Andrej Karpathy’s char-rnn) following Jeff Thompson’s instructions. I then got a list of all the Pokemon names from all generations from Bulbapedia. I had to write a simple script to parse out the names from the article’s source before feeding it into the preprocessor.

Then, I just fed it into the RNN with default settings and the start text “Rapidash” (a real Pokemon) and let it do its thing. You can find a list of a couple hundred generated Pokemon names here.

Some of my favorites:

Goomoonsher Charmant Dilolipch Bariboo Literona Gmoople Mingus Seeflip

Actually, I can’t tell if some of these are real names! I’m not familiar with many Pokemon after Generation I.

I would like to put this up as a web app somewhere but I don’t know where I could install torch easily on a free-tier server.

Up next: downloading images from Bulbapedia and trying to use some neural networks to generate random Pokemon art!