puder — Predicting User Drafts by Empirical Learning

This is a transformer-based model to predict picks in MTG drafts. For a detailed description, please see the article.

There is also an online version, which is the easiest way to run the model.

Running it locally

(If you want to do training, please see below.)

You need Python (I used Python 3.12.8, other recent version should also work). You don't need a GPU (unless you want to do training). This model is tiny and can easily do inference on the CPU.

You also need the source code, model weights, and some data. The easiest way to get all of that is to download a release.

Then do something like the following:

python -m venv .venv
. .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements
python draftpredict.py webui

Or, if you prefer uv, instead do

uv venv
. .venv/bin/activate
pip install torch --index-url https://download.pytorch.org/whl/cpu
uv pip install -r requirements
python draftpredict.py webui

You should now be able to access the UI at http://localhost:31180/ .

Training

Please beware, this is not very usable outside of my machine and you will likely have to reverse-engineer and fix some things.

• You probably want a GPU.
• You then need a version of PyTorch with hardware acceleration. See here for instructions.
• Download the public 17Lands datasets, and convert them to tensors using csv_to_tensor.cpp. You'll need philib and libdeflate to compile it.
• Install the other required packages in the same way as for inference.
• Runnning python draftpredict.py train should then create and train a new model.