Magic Autocomplete

This is a Magic: the Gathering decklist autocompletion tool

Made by Julien Audet, not affiliated in any way to Magic: the Gathering nor to Wizards of the Coast

This project was made as my CEGEP integrator project. It has been made over the course of 20 weeks and 100+ hours of research and work. A paper detailing the development process and results of the first version is available at: https://docs.google.com/document/d/1LHNd_-JQrEzRyDKt-MnAX0Gg7qHPufJPXdoyU1dHDbQ/edit?usp=sharing.

This project aims to predict Magic: the Gathering decklists from subsets of those lists, as a player would do when facing an opponent. A demo of the working tool is available on https://drive.google.com/file/d/1zIPOR6mSmvFgz_L1QG9-KxZphI7PVPhe/view?usp=sharing

Usage

The project is divided in five main parts: data acquisition (scraping), data preprocessing, model training, model evaluation, and model usage.

Setup

From the project directory, create and activate a Python venv:

# Linux 
python3 -m venv venv
source venv/bin/activate

# Windows
python -m venv venv
./venv/Scripts/activate

Then install the required dependencies:

pip install -r requirements.txt

You are now ready to use this project!

Data Acquisition

To minimize drift, up-to-date data should be scraped. For a simple demo of the autocompletion tool, you can skip ahead to the 'model usage' section. Raw data takes the form of text decklist like those available on MTGGoldfish.

The scraper code is found in data_preprocessing/scraper.py. First, ensure that each of the links to the different archetypes are up-to-date. You may add archetypes to the dictionary if you want. Update the paths found in data_preprocessing/config.py to match your environment. You can now start the scraper by running the data_preprocessing/scraper.py file. The code gathers one list per second to avoid overloading the website, so scraping 1000+ lists may take a while.

Here is an example of the scraper in action:

Data Formatting

Before encoding the data and training models, we need to format the data. A specific description of what this entails can be found in the project paper. To run data formatting, simply run the data_preprocessing/formatter.py file.

Examples of unformatted and formatted lists can be found in data_18_12_2023/.

Data Encoding

Some of the LSTM models take a card vector as input. This vector is found using the Word2Vec algorithm, which associates similar vectors to tokens found in similar contexts. More information about this method is given in the project paper.

To encode the cards into vectors, change the output Word2Vec model's name in data_preprocessing/vectorizer.py, and then run the file.

Card Vector Demo

A short Jupyter notebook demonstrates the vector's capability to encapsulate similarities between cards, as it does with words. Find this notebook in demo/card_vector_demo.ipynb. To use the demo, simply follow the instructions provided in the notebook. A video showcasing this notebook is available on https://drive.google.com/file/d/15rqoWeKmwyorjKtAdXo68dlZ5Bj_9X4E/view?usp=sharing

Model Training and Evaluation

The code for the original report, finished in May 2022, is available in original_project_files_2022.zip. Unfortunately, upon refactoring the project in January 2024, the improved model code made during the summer of 2022 had issues with the TensorFlow library, most likely due to the newer version used. Due to a lack of time on my end, I did not fix these issues, and so an up-to-date model training and evaluation script is not currently available. The most recent code for training, evaluating, and using the improved model is available in commit cb1699945fcba7d764c4acbda2a354bdf0b48567 with tag 'old_and_final_model_code', in directory final_model/.

Old models from the report are currently available in the models/lstm_2022/ directory. The meaning behind the names are explained in the project paper. For demo purposes, I recommend using L1000S.h5, as it gives good results. Be aware that these models were trained with decklists from January 2022, and so they might fail if new cards from recent competitive lists are used. For demo purposes, all the scraped lists from January 2022 are included in data_21_01_2022.

Model Usage

The code in demo/report_model_demo allows to complete decklists using the models trained for the project paper. To obtain predictions along with an accuracy measure, update the list in demo/report_model_demo/target_decklist.txt with a formatted decklist of your choice. Then, include a subset of this list in demo/report_model_demo/decklist_subset.txt. I recommend you include 4-10 cards in the subset, as it showcases the model's ability to capture the essence of a specific strategy. Update the paths found at the top of the demo/report_model_demo/report_model_demo.py file. You can now run the file to obtain predictions.

A video showcasing the predictions can be found on https://drive.google.com/file/d/1zIPOR6mSmvFgz_L1QG9-KxZphI7PVPhe/view?usp=sharing , as mentioned above.

If you like the project or find something that could be made better, please let me know! I'd love to hear what you think!