evoker

A webserver for serving Renderflux's GPT-2 based generative art prompt autocomplete model

Training

training is very simple. Just use the already provided prompt dataset located in /evoker/data/prompts.txt, or replace it with your own.

The format is just prompts, seperated by newlines, like so:

A summer field, art by tyler edlin, trending on artstation
a majestic dragon flying through the clouds by typer edlin, trending on artstation
a mountain with an orb of energy above it by typer edlin, trending on artstation
A dark and dreary place, where the only light is the moonlight, the only sound is the wind, and the only smell is the stench of decay

once you have your dataset, you can begin training. From the root directory of the project, install dependencies with poetry

poetry install

then enter the virtualenv

poetry shell

now start training!

python evoker train {epochs}

a good amount for epochs is 50-200. You can experiement with this though.

after training, you can test your model quickly like so:

python evoker predict {prompt}

for axample,

python evoker predict a beautiful

['a beautiful dreary place and charming mountain goat palace, great outdoors field, charming and often secluded']

Serving

once you have a good model trained up, you can start the webserver with

python evoker serve

once the webserver starts, you can make a POST request to /predict with a json body that looks something like:

{
    "prompt": "a beautiful",
    "amount": 1
}

thats pretty much it. If you have any issue don't hesistate to make a PR!