Is this available as a patch on a JIRA in addition to your web hosted source code? I think that this is very close to ready to commit. The missing test can be handled pretty easily because svd(A') = V D U' so all that is needed is to detect the problematic shape, decompose the transposed matrix and reverse the roles of U and V.
On Wed, Jul 21, 2010 at 1:09 AM, Nicolas Maillot <[email protected]> wrote: > Ted, Robin, > > Thanks for your remarks. I have taken them into account. > > I have added a set of unit tests for the SVD (based on > org.apache.commons.math.linear.SingularValueDecompositionImplTest): > > http://www.nicolas-maillot.net/MahoutContribs/TestSingularValueDecomposition.java > Note that I had to deactivate some tests due to limitations of the > CERN implementation: > -Tests where the input matrix has strictly fewer rows than columns > -Tests on svd covariance which is not (yet) implemented > > Otherwise all the "valid" tests are running fine. > > I have also fixed the parameter management. > > Nicolas > > > > > > On Mon, Jul 19, 2010 at 11:08 PM, Ted Dunning <[email protected]> > wrote: > > Nicolas, > > > > This was quick work. Very nice to see. > > > > To finish, the migration, we need to add test cases as well. > > > > A good source of test cases for the SingularValueDecomposition would > > be package > org.apache.commons.math.linear.SingularValueDecompositionImplTest > > > > On Mon, Jul 19, 2010 at 1:08 PM, Nicolas Maillot <[email protected]> > wrote: > > > >> As you can see, I have migrated SingularValueDecomposition (useful to > >> tackle the inverse covariance problem) to use the Matrix Class. > >> > > >
