Hello,
Thank you for your comments. I had actually initially tried your first
suggestion, but the predicted values just wouldn’t line up between the two
models. As I dug into the source code of the two, I realized that they don’t
appear to be the same. MLPY adds a bias term to both the training and
prediction process, whereas, correct me if I’m wrong, scikit-learn does not.
This results in two fundamentally different sets of codes:
MLPY (prediction):
np.dot(self._alpha, Kt_arr.T) + self._b
Scikit-learn(prediction):
np.dot(K, self.dual_coef_)
Here, MLPY’s alphas correspond to scikit-learn’s dual_coef_, and the kernel
values are just stored differently, so one has to be transposed. If I just try
and add MLPY’s bias term to scikit-learn’s prediction (model.predict), the
values don’t match those predicted by MLPY (they’re close but they are not off
by a constant value). Am I missing something obvious, or is there really a
fundamental difference here?
From: scikit-learn <scikit-learn-bounces+jchapman48=gatech....@python.org> on
behalf of Michael Eickenberg <michael.eickenb...@gmail.com>
Reply-To: Scikit-learn mailing list <scikit-learn@python.org>
Date: Thursday, September 19, 2019 at 3:53 PM
To: Scikit-learn mailing list <scikit-learn@python.org>
Subject: Re: [scikit-learn] Porting old MLPY KRR model to scikit-learn
What exactly do you mean by "port"? Put already fitted models into a sklearn
estimator object? You can do this as follows:
You should be able to create a `estimator =
sklearn.kernel_ridge.KernelRidge(...)` object, call `fit` to some random data
of the appropriate shape, and then set `estimator.dual_coef_` to the ones from
your MLPY model (the sklearn version sets them here:
https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/kernel_ridge.py#L165).
If this is not what you mean, then maybe you just want to refit them using the
appropriate KernelRidge kernel?
Hope this helps!
Michael
On Thu, Sep 19, 2019 at 12:43 PM Chapman, James E
<jchapma...@gatech.edu<mailto:jchapma...@gatech.edu>> wrote:
Hello,
I have some old KRR models from MLPY and I need to port those models over to a
new code written with scikit-learn (transfer MLPY KRR data to a scikit-learn
KernelRidge instance). Does anyone know if this is even possible, and if so,
could you give me some suggestions as to how to accomplish it?
Thanks and regards,
James
_______________________________________________
scikit-learn mailing list
scikit-learn@python.org<mailto:scikit-learn@python.org>
https://mail.python.org/mailman/listinfo/scikit-learn
_______________________________________________
scikit-learn mailing list
scikit-learn@python.org
https://mail.python.org/mailman/listinfo/scikit-learn
