Awesome

LKIS

This is a demo implementation of the following paper.

Naoya Takeishi, Yoshinobu Kawahara, and Takehisa Yairi, "Learning Koopman Invariant Subspaces for Dynamic Mode Decomposition," in Advances in Neural Information Processing Systems (Proc. of NIPS), vol. 30, pp. 1130-1140, 2017.

arXiv preprint: https://arxiv.org/abs/1710.04340

Prerequisite

• python 3.5.2 or later
• numpy 1.12.1 or later
• scipy 0.19.0 or later
• chainer 1.23.0

Files

• lkis.py
  • Core implementation of LKIS network.
• train.py
  • Script for training network.
• predict.py
  • Script for test by prediction based on a trained model.
• exp_lorenz
  • Root directory for experiment using Lorenz series. Dataset is included here.
• matlab/
  • MATLAB tools.

Usage

python train.py [name] [options]
python predict.py [name] [options]

[name] specifies the name of the experiment.

Example

python train.py lorenz --numval 1 --delay 7 --dimobs 5
python predict.py lorenz --save

The result can be inspected using matlab/exp_lorenz.m

Important options

train.py

• --rootdir
  • Root directory of an experiment. Data and results must be stored under this directory.
• --datadir
  • Name of the directory (under the root directory of the experiment) that contains datasets.
• --outputdir
  • Name of the directory (under the root directory of the experiment) where results will be stored.
• --numtrain
  • Number of training dataset files. If not specified, only one dataset file train.txt is used. If specified with 2, for example, train_0.txt and train_1.txt are used.
• --numval
  • Number of validation dataset files. If not specified, no validation dataset is used. If specified with 1, val.txt is used. If specified with 2, for example, val_0.txt and val_1.txt are used.
• --delay
  • Dimensionality of delay coordinates. $k$ in the paper.
• --dimemb
  • Dimensionality of the delay embedding. $p$ in the paper.
• --dimobs
  • Dimensionality of the learned observable. $n$ in the paper.
• --epoch
  • Number of epochs for SGD.

predict.py

• --numtest
  • Similar to --numval.
• --horizon
  • Timestep horizon to which prediction is calculated.
• --save

Author

• Naoya Takeishi - https://ntake.jp/

License

This project is licensed under the MIT License - see the LICENSE.txt file for details