Dynamics-Aware Models

Code for reproducing Physical Reasoning Using Dynamics-Aware Models. This branch trains models that use the hand crafted loss.

Getting Started

Installation

A Conda virtual enviroment is provided contianing all necessary dependencies.

git clone https://github.com/facebookresearch/DynamicsAware
cd DynamicsAware
conda env create -f env.yml
source activate dynamics_aware
# You might need to replace next command with the correct
# command to install pytorch on your system if you are not running linux
# and cuda 10.2  
conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch
pip install -e src/python

Running Experiments

To run an experiments locally

cd agents
python  python run_sweep_file.py  <experiment_file>  --base-dir=<basedir> -o -l

Where <experiment_file> is the experiment file to run and <basedir> is the directory where the experiment output should be stored. For more details see agents

License

Dynamics-Aware Models is released under the Apache license. See LICENSE for additional details.

Citation

If you use DynamicsAware or the baseline results, please cite it

@inproceedings{ahmed2021physical,
  title={Physical Reasoning Using Dynamics-Aware Models},
  author={Ahmed, Eltayeb and Bakhtin, Anton and van der Maaten, Laurens and Girdhar, Rohit},
  booktitle={ICML Workshop},
  year={2021}
}