Learning to Correspond Dynamical Systems

This is a supplementary codebase for Learning to Correspond Dynamical Systems (L2CDS). More details can be found on the project webpage.

Requirements

This software package has been tested on the following environment:

Ubuntu 18.04 LTS
DART v6.6.2
Python 3.7 or above

Following Python packages are required (not comprehensive):

scikit-learn==0.22.0 (importing pre-computed results will break otherwise)
PyTorch (torch==1.9.1+cu111 recommended)
PyDart2
matplotlib
tqdm
stable-baselines3

With Docker / Singularity

NOTE: Visualization with GLUT window isn't really supported inside containers. For visualization, I recommend using the non-container route, although it involves more headache.

A Dockerfile is provided for your convenience. The pre-built image is found on Docker Hub: https://hub.docker.com/r/namheegordonkim/l2cds

If you use Docker, use

docker pull namheegordonkim/l2cds

After pulling the latest l2cds image, please mount the local repo directory to /l2cds within the container, i.e.

docker run -it --mount type=bind,source=$(pwd),target=/l2cds namheegordonkim/l2cds

If you use Singularity, use

singularity pull docker://namheegordonkim/l2cds

Without Docker / Singularity

Ensure that your OS-level Python is version 3.7 or above.
Follow the instructions here to build and install DART v6.6.2. That is, ensure to run git checkout tags/v6.6.2 before building with cmake and make.
Build PyDart2 from source, i.e. do something like:

git clone https://github.com/sehoonha/pydart2.git
cd pydart2
python setup.py build build_ext
python setup.py develop

NOTE: The way relative import works in Python 3.7+ seems to mess with the way SWIG generates bindings. In case any error regarding _pydart2_api shows up, do the following

Locate pydart2/pydart2.api.py
Change relative import statement from . import _pydart2_api to import _pydart2_api.

Install dart-env. You will want to build it from source in develop mode using pip install -e ".[dart]", i.e.

git clone git clone https://github.com/DartEnv/dart-env.git
cd dart-env
pip install -e ".[dart]"

Install the rest of the requirements.

Running Experiments

When running within the docker container, use python3. If not, use python3 or python, whichever works.
The training scripts will dump snapshots at a specified interval (save_every in config files), and these will be used for computing correspondence.
By default, the config files specify experiments to run for large numbers of iterations. However, you can babysit for the loss value and terminate the experiment whenever, and use the latest dumped artifacts.

Shapes Experiments

Make shapes

python shapes_simulate.py

Correspond the shapes

python shapes_learn_correspondence.py

Visualize correspondence

python shapes_visualize.py

Walker Experiments

Coming soon.

namheegordonkim / l2cds