This repository contains source code for the following papers:
- ICML 2024: On the identifiability of Switching Dynamical Systems.
- CI4TS Workshop @ UAI 2024: Identifying Nonstationary Causal Structures with High-Order Markov Switching Models .
The source code is build on Python 3.8.19 and the models are implemented using Pytorch 2.3. Run pip install -r requirements.txt to install the dependencies.
Below we provide simple indications to test the code.
Generate some synthetic data (defaults to a 2D MSM with 3 states)
python generate_data_and_train_msm.py --generate_data --no-train --seq_length 100 --num_samples 10000
Run python train_msm.py.
Generate some synthetic data (defaults to a 2D SDS with 2D latent MSM and 3 states)
python generate_data_and_train_snlds.py --generate_data --no-train --seq_length 200 --num_samples 5000
Run python train_snlds.py > log_train_SDS.log. It is convenient to write the outputs in a separate text file, e.g. log_train_SDS.log.
To reproduce the results, the files generate_data_and_train_msm.py and generate_data_and_train_snlds.py can be used.
For example, the following command trains the models to generate the lower left cell of Figure 4b in our paper.
python generate_data_and_train_msm.py --seeds 23 24 25 26 27 --dim_obs 3 --num_states 3 --sparsity_prob 0.0 --data_type cosine --device cuda:0 --generate_data --restarts_num 20 --seq_length 200 --num_samples 10000 > log_train_3_states_3_dim.log
The logs can be found in log_train_3_states_3_dim.log. For other settings like polynomials or softplus networks, the --data_type argument can be set to poly or softplus respectively. poly defaults to cubic polynomials used in the paper, use --degreee to change the degree type.
The following command trains a Switching Dynamical System on images using different seeds, which can be used to generate the rightmost column of Figure 6 in our paper.
python generate_data_and_train_snlds.py --seeds 23 24 25 26 27 --dim_obs 2 --images --dim_latent 2 --num_states 3 --sparsity_prob 0.0 --data_type cosine --device cuda:0 --generate_data --restarts_num 5 --seq_length 200 --num_samples 5000 > log_train_SDS_images.log
The logs can be found in log_train_SDS_images.log, and the models will be saved in results/models_sds. Note the script can take a very long time due to the number of seeds and restarts used. Consider using more restarts for optimal results.
For evaluation scripts and trained models, contact the corresponding author.
If you use this code, or you find our paper useful for your research, consider citing our paper.
@inproceedings{
balsells-rodas2024on,
title={On the Identifiability of Switching Dynamical Systems},
author={Carles Balsells-Rodas and Yixin Wang and Yingzhen Li},
booktitle={Forty-first International Conference on Machine Learning},
year={2024}
}