This repository contains the code for our CDC2024 submission: Jing Cheng, Ruigang Wang, and Ian R. Manchester, Learning Stable and Passive Neural Differential Equations.
This code has been tested with Python 3.8.10. The figures are generated with MATLAB R2022b.
We assume any version later than Python 3.8.10 should work, also please install the packages listed in requirements.txt
To train the models, run following script to train and save the models:
train.py
The models will be saved in ./results/pendulum2/ dir.
Then the simulation error is calculated via:
simulation_error.py
The trjectory of physical model will be at ./dataset/p-physics-2.npy, and all the results from neural dynamics are in ./results/simu_error/
-
loss_trainsize_plot.mgives the loss v.s. training size figure, data from training with diffrerent training sizes. -
mu_nu_tuning.mgives the loss v.s.$\mu / \nu$ figure, data from training with different parameters$\mu$ and$\nu$ -
loss_plot.mgives the learning loss figure -
trajectory_plot4.mgives the simulation trajectory (note: the trajectory figure in the paper was plot with two close initial conditions to better illustrate our strength. Here we just show two batches) -
error_plot.mgives the simulation error figure
Run the follow to test a few training sizes:
./train_check_trainsize
For any questions or bugs, please raise an issue or contact Ruigang (Ray) Wang (ruigang.wang@sydney.edu.au) or Jing (Johnny) Cheng (jing.cheng@sydney.edu.au)