Awesome

TypeT5: Seq2seq Type Inference using Static Analysis

This repo contains the source code for the paper TypeT5: Seq2seq Type Inference using Static Analysis.

@inproceedings{Wei2023TypeT5,
    title={TypeT5: Seq2seq Type Inference using Static Analysis},
    author={Jiayi Wei and Greg Durrett and Isil Dillig},
    booktitle={International Conference on Learning Representations},
    year={2023},
    url={https://openreview.net/forum?id=4TyNEhI2GdN}
}

Installation

This project uses pipenv to manage the package dependencies. Pipenv tracks the exact package versions and manages the (project-specific) virtual environment for you. To install all dependencies, make sure you have pipenv and Python 3.10 installed, then, at the project root, run the following two commands:

pipenv --python <path-to-your-python-3.10>  # create a new environment for this project
pipenv sync --dev # install all specificed dependencies

More about pipenv:

• To add new dependences into the virtual environment, you can either add them via pipenv install .. (using pipenv) or pipenv run pip install .. (using pip from within the virtual environment).
• If your pytorch installation is not working properly, you might need to reinstall it via the pipenv run pip install approach rather than pipenv install.
• All .py scripts below can be run via pipenv run python <script-name.py>. For .ipynb notebooks, make sure you select the pipenv environment as the kernel. You can run all unit tests by running pipenv run pytest at the project root.

If you are not using pipenv:

• Make sure to add the environment variables in the .env file to your shell environment when you run the scripts (needed by the parsing library).
• We also provided a requirements.txt file for you to install the dependencies via pip install -r requirements.txt.

Using the trained model

The notebook scripts/run_typet5.ipynb shows you how to download the TypeT5 model from Huggingface and then use it to make type predictions for a specified codebase.

Training a New Model

• First, run the notebook scripts/collect_dataset.ipynb to download and split the BetterTypes4Py dataset used in our paper.
  • The exact list of repos we used for the experiments in paper can be loaded from data/repos_split.pkl using pickle.load. They can also be downloaded via this Google Drive link.
• Then, run scripts/train_model.py to train a new TypeT5 model. Training takes about 11 hours on a single Quadro RTX 8000 GPU with 48GB memory.

Development

• Formatter: We use black for formatting with the default options.
• Type Checker: We use Pylance to type check this codebase. It's the built-in type checker shipped with the VSCode Python extension and can be enabled by setting Python > Anlaysis > Type Checking Mode to basic.