PriorKnowledgeNeuralODE

C. Coelho, M. F. P. Costa, and L.L. Ferrás, "A Study on Adaptive Penalty Functions in Neural ODEs for Real Systems Modeling" in Proceedings of the International Conference of Numerical Analysis and Applied Mathematics (ICNAAM-2023) (AIP Conference Proceedings, accepted)

This library provides a torch implementations of 3 adaptive penalty functions that can be used for training NN architectures. To learn more check the paper.

Installation

pip install PriorKnowledgeNeuralODE

Dependencies

1. torchdiffeqq
2. torch
3. pandas
4. numpy
5. math
6. matplotlib

Examples

There are 2 case study examples that use a Neural ODE to model the World Population Growth and the evolution of a Chemical Reaction available here

Options

1. --method :numerical method to solve the ODE, choices=['dopri5', 'adams']
2. --data_size :number of training time steps/li>
3. --test_data_size :number of testing time steps
4. --niters :number of iterations to train the NN
5. --test_freq :frequency to compute and print the test metrics
6. --gpu :turn on/off gpu
7. --adjoint :use the adjoint method to compute the gradients
8. --tf :value of the last time step for training
9. tf_test :value of the last time step for testing
10. --savePlot :path to store the plot of the real vs predicted curves
11. --saveModel :path to store the weights of the trained model
12. --adaptiveFunc :choice of the adaptive penalty function choices=['self', 'lemonge', 'dynamic0', 'dynamic1']

If you found this resource useful in your research, please consider citing.

@inproceedings{,
  title={A Study on Adaptive Penalty Functions in Neural ODEs for Real Systems Modeling},
  author={Coelho, C. and Costa, M. F. P. and Ferrás, L. L.},
  journal={International Conference of Numerical Analysis and Applied Mathematics (accepted)},
  year={2023}
}