Karel Synthesis

This work-in-progress repository implements a symbolic DSL for Karel and implements the Program VAE method from LEAPS.

Instalation

Download the LEAPS dataset from this link, and extract on data/program_dataset. The file tree should look like this:

data
└── program_dataset
    ├── data.hdf5
    └── id.txt

Install the requirements from PyPI:

pip install -r requirements.txt

Program VAE implementation

The Program VAE models are implemented in vae/models. base_vae.BaseVAE contains the base implementation of common functions and modules for Program VAEs and leaps_vae.LeapsVAE implements the Program VAE from the LEAPS paper.

Program dataset pre-loading

In order to speed up the trainer, the script main_dataset.py can be used to pre-load the dataset to a pickle file:

python3 main_dataset.py

Training Program VAE

LeapsVAE can be trained by running main_trainer.py:

python3 main_trainer.py

This command instantiates a Trainer from vae/trainer.py and runs it directly.

For information about the available arguments, run the script command:

python3 main_trainer.py --help

Latent Search

Also known as CEM in the LEAPS paper, Latent Search is implemented in search/latent_search.py. The script main_latent_search.py contains a wrapper to that class and can be called directly to execute Latent Search. In order to get the same behaviour as the original LEAPS paper, execute it with the following arguments:

python3 main_latent_search.py --env_enable_leaps_behaviour --search_reduce_to_mean

The available arguments are the same as the trainer script, and can similarly be visualized through:

python3 main_latent_search.py --help