gpt-mtg

This repository holds the code for generating mtg cards with gpt2.

Getting Started

Build Containers

To build containers capable of running the code on either the cpu or gpu use the command make build-cpu or make build-gpu.

Finetuning Model

To finetune a model on your dataset:

1. Create a dataset like output.txt in the wonderful mtgencode repository. If you want to build your own dataset follow the instructions in the mtgencode repostiory. For this project I just used output.txt
2. Create the appropriate docker container by running either make build-cpu or make build-gpu
3. Run python3 run.py finetune with the appropriate flags, with no adjustments this will run on the magic data in the data folder.

Predicting Text

To predict text:

1. Download the models into the gpt-mtg base directory from here
2. Unzip the files, the gpt model should unzip into a folder called checkpoint
3. open up a docker container with make run-cpu
4. run the text generation script python3 run.py generate-text

Creating Cards

Follow the instructions in this repository to get setup. Once you have all of the necessary tools to generate magic cards then:

1. Start docker with make run-cpu
2. Run the model to generate an encoding which should be parsable by mtgencode by running python3 run.py generate-text --destination-path <filepath.txt>
3. Use mtgencode to create a human readible version of the card ./decode.py -v <cardfile.txt> cards.pretty.txt -d
4. Use mtgencode to create a magic set zeditor set filedecode.py -e rfields -mse encoded.txt MSE/card
5. Use magic set editor to create sets from the cards

Contributing

To contribute to this repository please:

1. cp dev/hooks/pre-commit .git/hooks and chmod +x .git/hooks/pre-commit

TODO:

• figure out how to get allsets.json and parse it using mtgencode
• figure out how to use decode to mse to get mse parseable cards
• Figure out how to use mse to generate new cards.