crestonbunch / neural-namer

Fantasy name generator in TensorFlow

Neural Namer

A character RNN that learns how to emulate the styles of names of different fantasy authors.

Demo: https://crestonbunch.github.io/neural-namer-demo/

Usage

Recommended process is to setup a Python Virtual environment.

First time setup

This will create a Python virtual environment, and install the necessary packages only for this project.

$ pip install virtualenv
$ virtualenv -p /usr/bin/python3 venv
$ source venv/bin/activate
$ pip install -r requirements.txt
$ python setup.py develop

Running from the virtual environment

Make sure you run the commands from inside the virtual environment, once the virtual environment is created, you can enter it with:

$ source venv/bin/activate

Scrape data

$ crawl wikia --out crawler/wikia/data/names.csv

Train the model

$ model train --data crawler/wikia/data/names.csv --save modeler/logs/1

Generate names

$ model gen --save modeler/logs/1 \
            --data crawler/wikia/data/names.csv \
            --author "Tolkien"

Replace 'Tolkien' with another author:

• Tolkien (Lord of the Rings)
• George Martin (A Song of Ice and Fire)
• Robert Jordan (Wheel of Time)
• Steven Erikson (Malazan)
• Brian Jacques (Redwall)
• Frank Herbert (Dune)
• Andrzej Sapkowski (The Witcher)

Web interface

TODO: currently only networks with one LSTM cell are supported by the web interface. Don't try to migrate a model with more than one LSTM cell! It might work, but your model certainly won't generate correct outputs.

To migrate Tensorflow models into the web directory:

$ python scripts/migrate.py \
    --data crawler/wikia/data/names.csv\
    --checkpoint modeler/logs/1/model.ckpt-5600

Replace the -5600 suffix with the last checkpoint in your directory

To run the web interface

$ cd web/
$ yarn install
$ webpack-dev-server
$ firefox localhost:8080

There is no web backend, which means you can serve the web interface from any service that can serve static HTML, CSS, and JavaScript. E.g. GitHub pages.