Unearthing InSights into Mars: unsupervised source separation with limited data

Code to partially reproduce results in Unearthing InSights into Mars: unsupervised source separation with limited data, published in the proceedings of ICML 2023.

Installation

Run the commands below to install the required packages.

git clone https://github.com/alisiahkoohi/srcsep
cd srcsep/
conda env create -f environment.yml
conda activate srcsep
pip install -e .

After the above steps, you can run the example scripts by just activating the environment, i.e., conda activate srcsep, the following times.

Scripts

Deglitching can be done for a toy example by running the following:

python scripts/toy_example.py

The default command line arguments are stored at configs/toy_example.json. Non-default arguments can be passed to the script by for example:

python scripts/toy_example.py
    --max_itr 1000 \
    --j 8,8 \
    --q 1,1 \
    --type exp_glitch

The generated data is stored in data/checkpoints/ directory. To visualize the results, run:

python scripts/visualize_results.py
    --max_itr 1000 \
    --j 8,8 \
    --q 1,1 \
    --type exp_glitch

The figures will be stored in the plots/ directory.

Note regarding caching: The scattering covariance computation caches the results in srcsep/_cached_dir and following runs with the same exact setup will simply load the results. Feel free to delete the cache when needed.

Questions

Please contact alisk@rice.edu for questions.

Authors

Rudy Morel and Ali Siahkoohi