hi.

updates

TITLE: TAKING A WALK WITH GPT-2

Here's a new project I'm working on!

Basically, I want to generate written descriptions of walks so that people can go on a walk via a website, since we are all stuck inside.

The core concept I'm adapting is the experience of reading a page in a choose-your-own-adventure story. Putting aside the mechanics of choose-your-own-adventure, I want to recreate the feeling of reading the short scene itself - a narrative designed to stand independent of whatever precedes or follows it.

Instead of giving the user choice (or at least, the illusion of choice), the user will only be able to proceed forward or end the "walk." Instead, descriptive scenes generated from a trained gpt2 model will be presented at random. The intent here is to mimic the spontaneity and unpredictability of a meandering walk, where one has no particular destination and instead simply walks for a period of time.

So, that's the overall concept! I'll probably finesse the mechanics when I get to the point of building the front end.

So far, my focus has been on preparing a corpus for gpt2 training. It took me a while to think of how I could amass text samples that modeled what I'm looking for, but after my nth hour googling "landscape description corpus" I realized that travelogues would provide something at least close to what I was hoping for.

This took me to Project Gutenberg, where I started with Audubon's diaries and ended up with a couple dozens similar travel … [READ MORE]