noisyoscillator / from_fourier_to_koopman

Linear and non-linear spectral forecasting algorithms

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

From Fourier to Koopman: Spectral methods for Long-term Forecasting

Paper: [From Fourier to Koopman: Spectral Methods for Long-term Time Series Prediction (jmlr.org)](https://jmlr.org/papers/v22/20-406.html)

Fourier and Koopman constitute spectral algorithms for learning linear and non-linear oscillators from data respectively. Both algorithms solve a global optimization problem in frequency domain and allow for modeling of systems of any dimensionality. The algorithms are written in numpy and pytorch.

The objective of the algorithms is, respectively:

Koopman and Fourier forecasting objectives

This code accompanies the following paper. A video abstract of the paper can be found here. Results on modeling a number of fluid flows can be found here.

-----------------How to use Fourier -----------------

Fourier fits a linear oscillator to data. The number of frequencies k that the signal is assumed to exhibit and a learning rate needs to be specified. It is recommended to whiten the signal (zero-mean and unit-variance).

To learn the oscillator from data, do:

from fourier_koopman import fourier
import numpy as np

x = (np.sin([2*np.pi/24*np.arange(5000)]) + np.sin([2*np.pi/33*np.arange(5000)])).T
x = x.astype(np.float32)

f = fourier(num_freqs=2)
f.fit(x[:3500], iterations = 1000)

To perform forecasting, do:

x_hat = f.predict(5000)

-----------------How to use Koopman -----------------

Because of the many different ways in which the Koopman algorithm can be utilized, running it is more involved and might require writing a custom model_object. Below we provide an example where f is a simple MLP and the squared loss is employed. In general, the class koopman is instantiated with a model object that specifies:

  • the topology of f
  • the loss
  • the number of frequencies num_freqs
from fourier_koopman import koopman, fully_connected_mse
import numpy as np

x = np.sin(2*np.pi/24*np.arange(5000))**17
x = np.expand_dims(x,-1).astype(np.float32)

k = koopman(fully_connected_mse(x_dim=1, num_freqs=1, n=512), device='cuda:0')
k.fit(x[:3500], iterations = 1000, interval = 100, verbose=True)

To perform forecasting, do:

x_hat = k.predict(5000)

Examples

In the following, a more involved example is given that uses a 1D tranpose-convolutional Neural Network to learn a traveling wave.

TO DO: ADD EXAMPLE

Koopman and Fourier forecasting objectives

License

Please see the LICENSE file.

About

Linear and non-linear spectral forecasting algorithms

License:GNU General Public License v3.0


Languages

Language:Jupyter Notebook 78.8%Language:Python 21.2%