time delays not working correctly about pykoopman HOT 7 CLOSED

It is because the pykoopman you imported is the one from PyPi, which is a very old version. I just uploaded the 1.0.1 version. You should be able to find the example working now.

from pykoopman.

Thanks! I uninstalled and reinstalled, tried both from git clone and pip install pykoopman.
But it for some reason is not loading correctly:
Traceback (most recent call last):

Cell In[891], line 1
EDMDc = pk.regression.DMDc()

AttributeError: module 'pykoopman' has no attribute 'regression'

I definitely have "regression" folder inside the git cloned folder...

EDIT: it works as from pykoopman.regression import DMDc, so I can do that.

from pykoopman.

please take a look at the cell [5] in https://github.com/dynamicslab/pykoopman/blob/master/docs/tutorial_koopman_hankel_dmdc_for_vdp_system.ipynb

So basically, if you are implementing EDMDc with time delay, set svd_output_rank or svd_rank less than the full rank might lead to worse performance.

In fact, this is not a special case. Here is an example in my JFM paper:

• Looking at figure 12 in the following paper, when using EDMD with monomials, choosing a svd rank in the least square does not help.

Reference

• Sparsity-promoting algorithms for the discovery of informative Koopman-invariant subspaces
  S Pan, N Arnold-Medabalimi, K Duraisamy
  Journal of Fluid Mechanics 917, A18

from pykoopman.

There seems to also be a problem instituting the SVD rank:

from pykoopman.regression import EDMDc
EDMDc = DMDc(svd_rank=20)

n_delays = 71
obs = TimeDelay(n_delays=n_delays)

model = pk.Koopman(observables=obs, regressor=EDMDc)
model.fit(X1.T, y=X2.T, u=U[:,n_delays:len(X)].T) 

gives the error:

model.fit(X1.T, y=X2.T, u=U[:,n_delays:len(X)].T) 
Traceback (most recent call last):

  Cell In[3541], line 1
    model.fit(X1.T, y=X2.T, u=U[:,n_delays:len(X)].T)

  File ~\anaconda3\envs\DMD\lib\site-packages\pykoopman\koopman.py:170 in fit
    self._pipeline.fit(x, y, regressor__u=u, regressor__dt=dt)

  File ~\anaconda3\envs\DMD\lib\site-packages\sklearn\pipeline.py:394 in fit
    self._final_estimator.fit(Xt, y, **fit_params_last_step)

  File ~\anaconda3\envs\DMD\lib\site-packages\pykoopman\regression\_base_ensemble.py:107 in fit
    self.regressor_.fit(X, y_trans, **fit_params)

  File ~\anaconda3\envs\DMD\lib\site-packages\pykoopman\regression\_dmdc.py:146 in fit
    self._fit_unknown_B(X1, X2, C, r, rout)

  File ~\anaconda3\envs\DMD\lib\site-packages\pykoopman\regression\_dmdc.py:175 in _fit_unknown_B
    assert rout <= r

AssertionError

It works fine when I don't include the svd_rank parameter.

from pykoopman.

can you have a complete self-contained code here? including the data

from pykoopman.

Thanks, that's really helpful!

from pykoopman.

So basically, if you are implementing EDMDc with time delay, set svd_output_rank or svd_rank less than the full rank might lead to worse performance.

It seems like when there is noise in the control signal, it unfortunately leads to overfitting

from pykoopman.