This is a package that uses BabyAI in a sender/receiver setup.
First, install this BabyAI fork and the corresponding MiniGrid fork, which ease compatibility. (If you’re feeling lucky, you can instead try installing the original BabyAI repo (and the corresponding original MiniGrid repo).)
Then, clone this repository and install it with pip3:
git clone https://github.com/thomasaunger/babyai_sr.git
cd babyai_sr
pip3 install --editable .
In order to use the plotting and visualization scripts as is, you’ll also need to install LaTeX.
This package is organized similarly to the original BabyAI repo. Models can be trained using scripts/train_rl.py.
The resulting models can be tested using scripts/test_rl.py, which also saves the episode data to an NPY file, which in turn can be visualized using the vis scripts.
The learning curves as described by the log files can be plotted using scripts/plot_curves.py.