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
- python 3.5.2 or later
- numpy 1.12.1 or later
- scipy 0.19.0 or later
- chainer 1.23.0
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.
python train.py [name] [options]
python predict.py [name] [options]
[name]
specifies the name of the experiment.
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
-
--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
andtrain_1.txt
are used.
- Number of training dataset files. If not specified, only one dataset file
-
--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
andval_1.txt
are used.
- Number of validation dataset files. If not specified, no validation dataset is used. If specified with 1,
-
--delay
- Dimensionality of delay coordinates.
$k$ in the paper.
- Dimensionality of delay coordinates.
-
--dimemb
- Dimensionality of the delay embedding.
$p$ in the paper.
- Dimensionality of the delay embedding.
-
--dimobs
- Dimensionality of the learned observable.
$n$ in the paper.
- Dimensionality of the learned observable.
-
--epoch
- Number of epochs for SGD.
--numtest
- Similar to
--numval
.
- Similar to
--horizon
- Timestep horizon to which prediction is calculated.
--save
- Naoya Takeishi - https://ntake.jp/
This project is licensed under the MIT License - see the LICENSE.txt file for details