An extended Echo State Network (ESN) for chaotic time series prediction and anomaly detection.
This is a new implementation of the framework used in my thesis.
If you are looking for the legacy torsk that was used there you
can find it here.
In addition to a randomly initialized input matrix this implementation makes it
possible to use convolutions, discrete fourier transforms, and gradients of images
as inputs to the ESN.
For a demonstration of the predictive power of the extended ESN check out the next three links that compare ESN, LSTM, and cycle-based of three different problems.
Prediction of a Gaussian blob that moves according to a Lissajous figure that is defined by:
x = sin(t)
y = cos(0.3*t)
The true evolution of the time series is visible in the top left, the ESN prediction in the top right, LSTM prediction lower left, and the cycle-based prediction in the lower right.
Prediction of a chaotically moving Gaussian blob. The trajectory of the maximum is governed by the Mackey-Glass time series in the x-dimension and a sine in the y-dimension. The true evolution of the time series is visible in the top left, the ESN prediction in the top right, LSTM prediction lower left, and the cycle-based prediction in the lower right.
Prediction of the Kuroshio region at the coast of Japan. Again: The true evolution of the time series is visible in the top left, the ESN prediction in the top right, LSTM prediction lower left, and the cycle-based prediction in the lower right.
Running with pip install -e . installs the torsk package which comes with some
convenience scripts to inspect prediction outputs.
After running one of the experiments you can analyse the output files
with one of the commands listed by running torsk --help
Switching from Numpy to PyTorch (and soon to Bohrium!) backends can be done by using the corresponding Numpy/Torch classes. An example usage which makes it possible to run the prediction with different backends like this:
python experiments/chaotic_lissajous/conv_run.py backend torch dtype float32
To run the tests, install torsk via pip install -e ".[test]"and run the
test by executing pytest in the main repo directory.
To see logging call during testing you can do:
pytest --log-cli-level=INFO
To run the tests with flake8:
pytest --flake8