PNLSS Code

This repo provides the necessary code for reproducing the results of the paper A projected nonlinear state-space model for forecasting time series signals (arxiv). If you are interested in only running the PNLSS model, please refer to this repository.

Setup

GPU is required to run the codes. The following docker image is built for CUDA12. For setting up the environment we assume that you have installed Docker on your machine. First clone the repository and build the image:

git clone https://github.com/christiando/pnlss.git
cd pnlss
git submodule update --init --recursive

docker build -t pnlss .

This might take a couple of minutes. Alternatively you can also use/pull the built image from dockerhub chdonner/pnlss. If not already installed, install NVIDIA Container Toolkit by following instructions and run in the following.

sudo systemctl restart docker

Once this is done you can start a jupyter lab environment.

docker run --gpus all -p 8888:8888 -v $(pwd):/app/work pnlss

Click on the link, that is prompted in your terminal, and you can execute the notebooks that come with this repository.

Running the codes for the figures

Figure 2: Run Fig2_PNLSS-VanderPol.ipynb

Figure 3: Run Fig3_PNLSS-Lorentz_computation-time.ipynb

Figure 4a: Run Fig4a_models_Dysts-Darts.ipynb

Figure 4b: Run Fig4b_models_correlations.ipynb

Figure 5: Run Fig5_violin_plot.ipynb

Figure 6: Run Fig6_combine_plots.ipynb

Figure 7: Run Fig7_vectorfield.ipynb

Running the codes for generating figure data

The repo comes already with the data for the figures. However, we also provide the code to reproduce the data. Be aware, that this can be time intensive. First start the docker image from within the pnlss folder:

docker run --gpus all -it -v "$(pwd):/app" pnlss /bin/bash

Figure 3:

Run the following

python notebooks/Fig3_computation_time_nlss.py
python notebooks/Fig3_computation_time_rbf.py
python notebooks/Fig3_computation_time_nlss.py --fixed
python notebooks/Fig3_computation_time_rbf.py --fixed

Figure 4b, 5, and 6: These figures use the fitted models and their prediction data. To re-run the benchmarking of the PNL-SS method only, run at pnlss/ (Each file takes ~3 days):

python repos/dysts/benchmarks/compute_benchmarks_noise_fine_high.py --pnlss_only
python repos/dysts/benchmarks/compute_benchmarks_noise_fine_low.py --pnlss_only

To compute the benchmarking of the PNL-SS and all the 13 methods in Darts, comment out the following lines in the codes and run the codes (This will take ~ 1 week):

python repos/dysts/benchmarks/compute_benchmarks_noise_fine_high.py
python repos/dysts/benchmarks/compute_benchmarks_noise_fine_low.py

The above codes use the optimized hyperparameters. To do the hyperparameter searchers of the PNL-SS and all the 13 methods in Darts, run (Each file takes ~ 3 weeks):

 repos/dysts/benchmarks/find_hyperparameters_high.py
python repos/dysts/benchmarks/find_hyperparameters_low.py

Figure 7:

Run the first file to compute the darts models for figure 7, and the second for the pnlss models.

python notebooks/Fig7_run_darts_models.py
python notebooks/Fig7_run_pnlss.py