Creative AI: Theory and Practice

This repository contains experimental code used for the KDL presentation, From Once Upon a Time, to Happily Ever After, via AI at the Creative AI: Theory and Practice symposium.

Set up

Install poetry and the requirements:

poetry install

Fine tune the text generation model

poetry run python creativeai/gpt.py MODEL_PATH DATA_PATH

• MODEL_PATH can either be a Hugging Face path or a local path.
• DATA_PATH should be a path to a text file with all the training data.

To see a list of all the available options for the text generation run the gpt.py script with the --help option:

poetry run python creativeai/gpt.py --help

usage: gpt.py [-h] [-n NAME] [-o OUTPUT] [-e EPOCHS] model data

Script to train a GPT model

positional arguments:
  model                 Model path
  data                  Text file with training data

optional arguments:
  -h, --help            show this help message and exit
  -n NAME, --name NAME  Model name
  -o OUTPUT, --output OUTPUT
                        Path to save the trained model
  -e EPOCHS, --epochs EPOCHS

Warning: Depending on the model being used, fine tunning a text generation model can be very time consuming without access to a GPU.

Run the notebook

The repository also contains a notebook with an interface to generate text and imgages from the generated text. The story document contains an example of a story generated in the notebook for the prompt: There was a man made of clockwork who longed to become human so, with the settings top p set to 0.9, and ngrams set to 4.

Warning: The notebook is set up to use GPUs.

Data

The models for the presentation were fine tuned using the Grimm's Fairy Tales.

Models

• GPT-Neo 1.3B for text generation;
• Stable diffusion for image generation.

Black, S., Leo, G., Wang, P., Leahy, C., & Biderman, S. (2021). GPT-Neo: Large Scale Autoregressive Language Modeling with Mesh-Tensorflow (1.0) [Computer software]. Zenodo. https://doi.org/10.5281/zenodo.5297715

Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., & others. (2020). The Pile: An 800GB Dataset of Diverse Text for Language Modeling. ArXiv Preprint ArXiv:2101.00027.

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., & Ommer, B. (2022). High-Resolution Image Synthesis With Latent Diffusion Models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 10684–10695.