🛶 Montlake

Welcome to Montlake. Montlake contains tools for geometric data analysis. It includes vector field group lasso and basis pursuit methods for parametric manifold learning. It also contains differentiable shape featurizations including interpoint distances, planar angles, and torsions from data positions for shape space analysis. It uses embedding and tangent estimation components of the megaman package.

Installation

You can install Montlake into Python directly from the command line without cloning the repository. We recommend the use of Anaconda.

conda create -n 'montlake' python=3.6
source activate montlake
conda install --channel=conda-forge -y pip nose coverage cython numpy seaborn scipy scikit-learn pyflann pyamg h5py plotly
pip install git+https://github.com/sjkoelle/montlake/
conda install nb_conda_kernels
python -m ipykernel install --user --name montlake --display-name "Python (montlake)”

Troubleshooting

You may also need to install libmagic (e.g. brew install libmagic), install pip (conda install pip), or install ipykernel (pip install ipykernel). If you are still having issues installing using pip, clone the repository (git clone git@github.com:sjkoelle/montlake.git) and run

cd montlake
nbdev_build_lib

If there is a compatibility issue with nb_conda_kernels, which may result in failure of installation of jupyter notebook, you can also try the following to install Montlake.

conda create -n 'montlake' python=3.6
source activate montlake
conda install pip
pip install ipykernel
python -m ipykernel install --user --name montlake --display-name "Python (montlake)”
conda install --channel=conda-forge -y nose coverage cython numpy scipy scikit-learn pyflann h5py plotly
pip install git+https://github.com/sjkoelle/montlake/
pip insall pyamg seaborn

Usage

After installation, clone the repository (git clone git@github.com:sjkoelle/montlake.git).

Real Data Experiments

Code from the experiments subfolder can be used recreate the MDS experiments results from Manifold Coordinates with Physical Meaning and Consistency of dictionary-based Manifold Learning.

Individual analyses are parameterized in experiments/configs and can be called from the command line. Please set ROOT_DIR to your Montlake path and DATA_DIR to a directory containing the data (which can be downloaded here). Please cite this paper if you use this data.

Please use the following commands to set paths.

export ROOT_DIR=yourmontlakepath
export DATA_DIR=yourdatapath

Subsequently, you may configure the .json files with appropriate local file system paths and run the following commands in terminal to reproduce the computational and plotting analyses of synthetic and real data experiments.

• Swiss Roll experiment:
  jupyter notebook experiments/swiss_roll_resubmission.ipynb

• Rigid Ethanol (no noise) Simulation:

jupyter notebook experiments/generate_rigid_ethanol_data.ipynb
python -u -m montlake.exec.run_exp --config $ROOT_DIR/experiments/configs/rigidethanol_diagram.json --outdir $DATA_DIR/rigidethanol --raw_data $DATA_DIR/raw_data/rigidethanol_nonoise.npy --nreps 25 --mflasso --name re_nonoise_diagram_mf
python -m montlake.exec.plotting --config $ROOT_DIR/experiments/configs_plotting/jmlr/reth_diagram_mf_jmlr.json

• Ethanol data with diagram dictionary

python -u -m montlake.exec.run_exp --config $ROOT_DIR/experiments/configs/ethanol_diagram.json --outdir $DATA_DIR/processed_data/eth_diagram_mflasso_122221 --raw_data $DATA_DIR/raw_data/ethanol.mat --nreps 25 --mflasso --name eth_diagram_mflasso_122221
python -m montlake.exec.plotting --config $ROOT_DIR/experiments/configs_plotting/jmlr/eth_diagram_mf_jmlr.json

• Ethanol data with full dictionary

python -u -m montlake.exec.run_exp --config $ROOT_DIR/experiments/configs/ethanol_full.json --outdir $DATA_DIR/processed_data/eth_full_mflasso_122221 --raw_data $DATA_DIR/raw_data/ethanol.mat --nreps 25 --mflasso --name eth_full_mflasso_122221
python -m montlake.exec.plotting --config $ROOT_DIR/experiments/configs_plotting/jmlr/eth_full_mf_jmlr.json

• Malonaldehyde data with diagram dictionary

python -u -m montlake.exec.run_exp --config $ROOT_DIR/experiments/configs/malonaldehyde_diagram.json --outdir $DATA_DIR/processed_data/mal_diagram_mflasso_122221 --raw_data $DATA_DIR/raw_data/malonaldehyde.mat --nreps 25 --mflasso --name mal_diagram_mflasso_122221
python -m montlake.exec.plotting --config $ROOT_DIR/experiments/configs_plotting/jmlr/mal_diagram_mf_jmlr.json

• Malonaldehyde data with full dictionary

python -u -m montlake.exec.run_exp --config $ROOT_DIR/experiments/configs/malonaldehyde_full.json --outdir $DATA_DIR/processed_data/mal_full_mflasso_122221 --raw_data $DATA_DIR/raw_data/malonaldehyde.mat --nreps 25 --mflasso --name mal_full_mflasso_122221
python -m montlake.exec.plotting --config $ROOT_DIR/experiments/configs_plotting/jmlr/mal_full_mf_jmlr.json

• Toluene data with diagram dictionary

python -u -m montlake.exec.run_exp --config $ROOT_DIR/experiments/configs/toluene_diagram.json --outdir $DATA_DIR/processed_data/tol_diagram_mflasso_122221 --raw_data $DATA_DIR/raw_data/toluene.mat --nreps 25 --mflasso --name tol_diagram_mflasso_122221
python -m montlake.exec.plotting --config $ROOT_DIR/experiments/configs_plotting/jmlr/tol_diagram_mf_jmlr.json

Code for the Tangent Space Lasso is also colocated in this repo.

Synthetic Data Experiments

Code from the synthetic_experiments subfolder can be used recreate the synthetic experiments results from Consistency of dictionary-based Manifold Learning. The data, results, and visualizations used in the paper can be found here.

Additionally:

• the code from the synthetic_experiments subfolder can be used to sample new data from the 3 synthetic manifolds described in the paper.
• rerun the experiments on the new sampled data and visualize the new results.
• use the sym_manifold package to create new manifolds which can be subsequently sampled.
• user friendly classes for running and configuring TSLasso and MLasso as well as container classes for the data and configuartions necessary for the two algorithms.

Contributing

Please feel free to contribute, branch, fork, or edit. This package was built using nbdev, and so source code is kept in the nbs folder and then compiled by running nbdev_build_lib. Please cite if you use!