to complexify a bit the pb note that in the SVM/Lasso/... case the
precomputed gram
is np.dot(X, X.T) which means that the cross-val can be done just with it
while for the covariance estimation, like GraphLassoCV, the empirical
covariance is np.dot(X.T, X) hence the fit needs X as input.

so it seems to me we have 3 cases:

- kernel / similarity, shape (n_samples, n_features)
- distance, shape (n_samples, n_features)
- cov, shape (n_features, n_features)

HTH,

Alex

On Wed, Nov 9, 2011 at 6:06 PM, Gael Varoquaux
<[email protected]> wrote:
> On Wed, Nov 09, 2011 at 11:43:40PM +0100, bthirion wrote:
>> > What do people think? Should I:
>
>> >   1. change graph_lasso to take the empirical covariance as an input
>
>> >   2. add an 'X_is_cov' parameter to the estimators
>> +1 for the second one.
>
> I actually was suggesting both, and 1 as a mean for 2.
>
>> If we want to introduce some kind of automated guess of the
>> regularization parameter, we'll have to know the dimension I believe ?
>
> You mean the number of samples? Actually, no, what is important is the
> number of degrees of freedom (I know that you know this). Things like the
> OAS try to estimate it from the covariance matrix.
>
> G
>
> ------------------------------------------------------------------------------
> RSA(R) Conference 2012
> Save $700 by Nov 18
> Register now
> http://p.sf.net/sfu/rsa-sfdev2dev1
> _______________________________________________
> Scikit-learn-general mailing list
> [email protected]
> https://lists.sourceforge.net/lists/listinfo/scikit-learn-general
>

------------------------------------------------------------------------------
RSA(R) Conference 2012
Save $700 by Nov 18
Register now
http://p.sf.net/sfu/rsa-sfdev2dev1
_______________________________________________
Scikit-learn-general mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/scikit-learn-general

Reply via email to