Maze Transformer

Warning

This repository is designed to be used with and . However, both libraries have been significantly updated since this repository was last maintained, and there may be compatibility issues between the two codebases if using latest versions.

Installation

pip install git+ssh://[email protected]/understanding-search/maze-transformer.git

Note: if you want to install the library in colab, follow the steps in this Colab notebook.

You can run all tests via

make test

(this will take a few minutes)

Usage

Most of the functionality is demonstrated in the ipython notebooks in the notebooks/ folder.

• demo_dataset.ipynb how to easily create a dataset of mazes, utilities for filtering the generates mazes via properties, and basic visualization
• train_model.ipynb configuration setup and training a basic model
• eval_model.ipynb loading a trained model, and computing various metrics on its performance on a dataset
• plot_attention.ipynb various attention visualization utilities
• demo_latticemaze.ipynb internals of the LatticeMaze and SolvedMaze objects, and advanced visualization
• train_model_hallway.ipynb training a model on a customized dataset

Development

Prerequisites

• Install Poetry

• Install Python 3.10

  • It's a good idea to use pyenv to manage python versions
  • If using pyenv, you'll need to update your Poetry config for it to use the pyenv Python version: poetry config virtualenvs.prefer-active-python true

• Install dev dependencies

  poetry config virtualenvs.in-project true
  poetry install --with dev
  

• Run unit, integration, and notebook tests

  make test
  

• (Optional) If you want to work with the jupyter notebooks in VSCode

  • create a jupyter kernel with poetry run ipython kernel install --user --name=maze-transformer
  • Restart VSCode
  • In VSCode, select the python interpreter located in maze-transformer/.venv/bin as your juptyer kernel

Testing & Static analysis

make help will print all available commands.

• all tests via make test

  • unit tests via make unit
  • integration tests via make integration
  • notebook tests via make test_notebooks

• formatter (black, pycln, and isort) via make format

  • formatter in check-only mode via make check-format

Cite As

@software{maze_transformer,
  author = {Michael I. Ivanitskiy and Alex F. Spies and Chris Mathwin and Dan Valentine and Can Rager and Guillaume Corlouer and Tilman Räuker and Rusheb Shah and Lucia Quirke},
  title = {Maze-Transformer: A Package to Train and Analyze Transformers on Maze-Solving Tasks},
  year = {2023},
  url = {https://github.com/understanding-search/maze-transformer},
}