Creative Wand

This repository holds the implementation of the Creative Wand framework, along with an exemplar implementation in text domain. For more details on what this is, feel free to check Creative Wand: A System to Study Effects of Communications in Co-Creative Settings by Zhiyu Lin, Rohan Agarwal, and Mark Riedl.

Installation & Setup

Use this as a library

• Create a conda environment using provided conda_env.yaml:
  • conda env create -f conda_env.yml
  • conda activate creative-wand
• pip install -e . and you will have CreativeWand.Framework available in your workspace.

Try the exemplar application

• Create a conda environment using provided conda_env.yaml:
  • conda env create -f conda_env.yml
  • conda activate creative-wand
• (Optional) Install apex library (Needed for loading models in half precisions (conserve memory): https://github.com/NVIDIA/apex
• Set up Addons
  • Clone https://github.com/eilab-gt/plug-and-blend-tool into Addons/PNB and rename the folder pnb2 (Or change endpoint.py accordingly)
  • Check AddonConfig.py and change paths accordingly
    • Download https://storage.googleapis.com/sfr-gedi-data/gedi_topic.zip , unzip it somewhere and point gedi_location there;
    • for base_location input a desired huggingface gpt2-compatible model (either a model name or a path to given model), or use gptj following the config sample.
    • We used gptj for our experiments, which requires a lot of VRAM. all functions work with any gpt-compatible models, although what is generated will be different.
• Additional information on Addons
  • Addons are individual modules fulfilling functionalities for our exemplar instantiation.
  • For your convenience, we built a simple REST API framework that can load these moduels, located at CreativeWand.Addons.
  • Check Sample.py for an example Addon.

Running the exemplar application

In scripts folder, run all scripts that start with start:

• start_addon_server.sh starts addon server;
• start_experiments_server.sh starts backend server;
• start_web_frontend_server.sh starts the Web interface server.
• Based on your system setup, you may need to chmod +x [script name] all scripts, or additionally deal with node.js dependencies.

Now you can access http://localhost:3000/?mode=mode_name&pid=test_session to start using the system.

• Available modes are local or global but check StartExperiment.py for how to add more (Customize available communications, Experience Manager, Frontend, etc.)
• pid will name the session, which then you can retrieve logs in logs/ folder.
• If you do not provide these URL parameters the start button won't work.

If it doesn't work

• If CreativeWand keeps saying it can't understand you
  • This is a handle for internal errors. Check the console to see whether there are configuration (connection, setup) issues between all backend servers.
• If Webserver complains about compile errors
  • npm install in web-frontend/web-interface (which should have been automatically done). There may be some per-machine node.js problem.
• If the Addon server gives error
  • Double check whether you have your Addons set up correctly. Did you correctly set the paths to the models downloaded?

Have fun!

Feel free to open an issue if you get into any problems :) .