Code for paper "Linear Dynamics: Clustering without identification"

Installation

The dependencies are documented in environment.yml and can be installed via conda.

conda env create -f environment.yml
conda activate lds

For more details, see conda documentation for creating an environment from yml file. If there is an error due to pybasicbayes and scipy version incompatibility, try installing the latest version of pybasicbayes from GitHub as the pip version might be outdated.

Simulated experiments for eigenvalue estimation

To run the eigenvalue estimation experiments in the paper, use the following commands.

python experiment_learn_eig.py --output_dir=eig_2d/ --min_seq_len=500 --max_seq_len=50000 --num_sampled_seq_len=5 --num_repeat=500 --hidden_dim=2 --output_noise_stddev=0.01
python experiment_learn_eig.py --output_dir=eig_3d/ --min_seq_len=500 --max_seq_len=50000 --num_sampled_seq_len=5 --num_repeat=500 --hidden_dim=3 --output_noise_stddev=0.01
python plot_eig_error.py

Simulated experiments for clustering

python experiments.py --output_dir=cluster_2d_2c --hidden_state_dim=2 --min_seq_len=1000 --max_seq_len=1000 --num_sampled_seq_len=1 --num_systems=100 --num_clusters=2 --num_repeat=100
python experiments.py --output_dir=cluster_2d_3c --hidden_state_dim=2 --min_seq_len=1000 --max_seq_len=1000 --num_sampled_seq_len=1 --num_systems=100 --num_clusters=3 --num_repeat=100
python experiments.py --output_dir=cluster_2d_5c --hidden_state_dim=2 --min_seq_len=1000 --max_seq_len=1000 --num_sampled_seq_len=1 --num_systems=100 --num_clusters=5 --num_repeat=100
python experiments.py --output_dir=cluster_2d_10c --hidden_state_dim=2 --min_seq_len=1000 --max_seq_len=1000 --num_sampled_seq_len=1 --num_systems=100 --num_clusters=10 --num_repeat=100
python agg_stats_to_table.py