Code release for Gamma-Models: Generative Temporal Difference Learning for Infinite-Horizon Prediction.
For the quickest startup, we recommend running the notebook directly in your browser using Google Colab.
This notebook will generate a video that looks like the following:
You can also use the trained model to perform value estimation, as shown in Figure 4 of the paper.
- Clone
gamma-models
git clone https://github.com/jannerm/gamma-models.git
- Create a conda environment and install
gamma
cd gamma-models
conda env create -f environment.yml
conda activate gamma
pip install -e .
- Add
gamma
as an IPython kernel and launch jupyter
python -m ipykernel install --user --name=gamma
jupyter notebook --port 6100 scripts
Open gamma-pendulum-local.ipynb
, which matches the Colab notebook except for a bit of Colab-specific setup in the beginning.
@inproceedings{janner2020gamma,
title={$\gamma$-Models: Generative Temporal Difference Learning for Infinite-Horizon Prediction},
author={Michael Janner and Igor Mordatch and Sergey Levine},
booktitle={Advances in Neural Information Processing Systems},
year={2020}
}
The underlying neural spline flow implementation is based on Andrej Karpathy's python-normalizing-flows repo, which in turn is based on Conor Durkan and Iain Murray's and nsf codebase.